Topics — The twenty-eight kinds of claim this anthology files against — the primary axis of both the practice registry and the reading list, and the same twenty-eight for each (ADR-026). The filing rule is the point: take a domain topic when the claim is ABOUT the domain; a claim merely discovered in one still takes its kind. The axis organizes the record, it does not bound it: content the list cannot place is evidence the list is short (ADR-059).
Generative modeling — diffusion, samplers, text-to-image, conditioning and control.
129 of 737 LIT documents. Back to the full index.
| # | Title | Summary | Status |
|---|---|---|---|
| LIT-036 v2 | Denoising Diffusion Probabilistic Models | Ho et al. (2020), ARXIV-2006.11239. Introduces DDPM framework. | Active |
| LIT-038 v3 | Denoising Diffusion Implicit Models | Song et al. (2020), ARXIV-2010.02502. Continuous-time formulation of diffusion models. | Active |
| LIT-062 v3 | High-Resolution Image Synthesis with Latent Diffusion Models | Rombach et al. (2022), ARXIV-2112.10752. Latent space diffusion. | Active |
| LIT-067 v3 | Progressive Distillation for Fast Sampling of Diffusion Models | Salimans et al. (2022), ARXIV-2202.00512. Faster sampling through distillation. | Active |
| LIT-070 v3 | Hierarchical Text-Conditional Image Generation with CLIP Latents | Ramesh et al. (2022), ARXIV-2204.06125. Prior-guided image generation. | Active |
| LIT-073 v2 | Photorealistic Text-to-Image Diffusion Models with Deep Language Understanding | Saharia et al. (2022), ARXIV-2205.11487. Cascaded diffusion approach. | Active |
| LIT-075 v2 | Elucidating the Design Space of Diffusion-Based Generative Models | Karras et al. (2022), ARXIV-2206.00364. Comprehensive design space analysis. | Active |
| LIT-076 v3 | DPM-Solver: A Fast ODE Solver for Diffusion Probabilistic Model Sampling in Around 10 Steps | Lu et al. (2022), ARXIV-2206.00927. Fast sampling for diffusion models. | Active |
| LIT-093 v3 | Consistency Models | Song et al. (2023), ARXIV-2303.01469. Single-step generation. | Active |
| LIT-097 v2 | One Transformer Fits All Distributions in Multi-Modal Diffusion at Scale | Bao et al. (2023), ARXIV-2303.06555. Universal sequence modeling. | Active |
| LIT-217 | Large Language Diffusion Models | Nie et al. (2025), ARXIV-2502.09992. LLaDA: an 8B language model trained from scratch under the ordinary pretrain-then-SFT paradigm, with the autoregressive factorization replaced by masked diffusion — a forward masking process and a reverse process that predicts masked tokens, optimizing a likelihood lower bound. Competitive with LLaMA3 8B on in-context learning, and beats GPT-4o on reversal poem completion. | Active |
| LIT-381 | Scaling Diffusion Language Models via Adaptation from Autoregressive Models | Gong et al. (2024), ARXIV-2410.17891. Diffusion language models are hard to train from scratch at scale and there are a great many trained autoregressive ones, so convert instead: continual pretraining on under 200B tokens turns GPT2 and LLaMA at 127M–7B into DiffuGPT and DiffuLLaMA. The conversion rests on a stated connection between the AR and diffusion objectives rather than on the two being retrained independently. | Active |
| LIT-439 | Deep Unsupervised Learning using Nonequilibrium Thermodynamics | Sohl-Dickstein et al. (2015), ARXIV-1503.03585. Where diffusion models come from: destroy structure with a slow forward process, learn the reverse, and get a generative model that is flexible and tractable — a trade the paper frames as the central problem it is solving. | Active |
| LIT-442 | Diffusion Beats Autoregressive in Data-Constrained Settings | Prabhudesai et al. (2025), ARXIV-2507.15857. Refits Muennighoff’s data-constrained scaling law for masked diffusion and gets a data-reuse half-life of 512 epochs against autoregressive training’s 32. Under a fixed corpus, AR wins at Chinchilla-optimal compute and then overfits; diffusion starts far worse and keeps improving to a lower final loss. A closed-form critical-compute threshold says which side of the crossover you are on. | Active |
| LIT-446 | Variational Diffusion Models | Kingma et al. (2021), ARXIV-2107.00630. Rewrites the diffusion VLB as an integral over signal-to-noise ratio, and the expression collapses: in continuous time the bound depends on the noise schedule only through its two endpoints. Variance-preserving and variance-exploding specifications are therefore equivalent up to a rescaling, and the schedule’s shape is free — spent here on minimizing the loss estimator’s variance. | Active |
| LIT-447 v2 | SiT: Exploring Flow and Diffusion-based Generative Models with Scalable Interpolant Transformers | Ma et al. (2024), ARXIV-2401.08740. Holds DiT’s architecture, parameter count and GFLOPs fixed and varies four transport choices one at a time: discrete vs continuous time, score vs velocity prediction, the interpolant connecting the distributions, and the sampler. Velocity prediction and a linear interpolant each help substantially; the sampler’s diffusion coefficient turns out to be choosable after training. | Active |
| LIT-448 | Scalable Diffusion Models with Transformers | Peebles and Xie (2022), ARXIV-2212.09748. Replaces the U-Net backbone every diffusion model had inherited from DDPM with a transformer over latent patches, and shows FID falls monotonically with forward-pass GFLOPs however the GFLOPs are bought — depth, width or token count. DiT-XL/2 reaches 2.27 FID on class-conditional ImageNet 256. | Active |
| LIT-449 v2 | Scaling Rectified Flow Transformers for High-Resolution Image Synthesis | Esser et al. (2024), ARXIV-2403.03206. A 61-way comparison of diffusion and rectified-flow formulations settles on a straight-line path with timesteps drawn from a logit-normal, then scales it to 8B for text-to-image. Adds a resolution-dependent shift of the timestep schedule, and MMDiT, which gives each modality its own weights while letting the two token streams attend jointly. | Active |
| LIT-477 | NoProp: Training Neural Networks without Full Back-propagation or Full Forward-propagation | Li, Teh and Pascanu (2025), ARXIV-2503.24322. Train each block independently to denoise a noisy label embedding given the raw input — diffusion machinery pointed at a classifier. No forward or backward pass across the network at training time, about half the GPU memory, and test accuracy matching backprop on MNIST and CIFAR-10. Every block sees the input, which is what makes independence possible and what makes the comparison awkward. | Active |
| LIT-479 v2 | The Diffusion Duality | Sahoo, Deschenaux, Gokaslan, Wang, Chiu and Kuleshov (2025), ARXIV-2506.10892. Uniform-state discrete diffusion is the argmax of a Gaussian diffusion, which lets Gaussian techniques transfer: a variance-reducing curriculum doubles training speed, and consistency distillation cuts sampling from 1024 steps to 8. Masked diffusion still wins on likelihood; this wins in the few-step regime, where masked models cannot revise what they have already emitted. | Active |
| LIT-483 | Mixture-of-Transformers: A Sparse and Scalable Architecture for Multi-Modal Foundation Models | Liang et al. (2024), ARXIV-2411.04996. Untie every non-embedding parameter by modality — feed-forward, attention projections, layer norms — and keep global self-attention over the whole sequence. Matches a dense baseline at 55.8% of the FLOPs in the Chameleon setting and 37.2% with speech added. The component ablation is the useful part: the feed-forward carries most of it, attention adds less, layer norms add nothing. | Active |
| LIT-488 | gen2seg: Generative Models Enable Generalizable Instance Segmentation | Khangaonkar and Pirsiavash (2025), ARXIV-2505.15263. Finetune Stable Diffusion or MAE end-to-end — encoder and decoder — for category-agnostic instance segmentation on indoor furnishings and cars only, and the model segments people, animals, x-rays and impressionist paintings. It approaches SAM across five domains on 0.3% of the annotated masks, and beats it ~3× on fine structures. The controls are the argument: the same backbone under a standard segmentation head scores near zero. | Active |
| LIT-490 | Denoising as Projection: Constrained Optimization with Gradient-Guided Diffusion | Zhang et al. (2026), ARXIV-2608.29507. Standard gradient guidance adds the objective gradient after the denoising step, which can push the sample off the learned data geometry. Swap the order — gradient first, then one pretrained-denoiser call — and the denoiser acts as an approximate projection, making the sampler an inexact projected-gradient method with convergence guarantees on linear subspaces, compact convex sets and compact Riemannian submanifolds. One gradient and one denoiser call per step; no score Jacobians, no fine-tuning, no extra samples. | Active |
| LIT-491 | Noise Hypernetworks: Amortizing Test-Time Compute in Diffusion Models | Eyring et al. (2025), ARXIV-2508.09968. Reward-guided test-time noise optimization costs 20–40 seconds a sample. Train a LoRA hypernetwork once to predict the optimized noise instead, and recover roughly half the gain at 0.1 s of added latency — 33–300× faster. The comparison that carries it: direct reward fine-tuning of the same generator makes it worse (GenEval 0.73 → 0.62 at four steps), because the KL term that would stop reward-hacking is intractable in data space and tractable in noise space. | Active |
| LIT-494 v2 | GaussianToken: An Effective Image Tokenizer with 2D Gaussian Splatting | Dong, Wang, Zheng, Chen, Lu and Tang (2025), ARXIV-2501.15619. Replaces the fixed grid of a VQ image tokenizer with K 2D Gaussians whose feature coefficient is quantized while position, rotation and scale stay continuous. Reports rFID 1.61 on ImageNet-1K against LlamaGen’s 2.19 with a 16× smaller codebook. The paper’s own word for the result is semi-discrete, and the comparison is at equal token count rather than equal rate — each token carries five floats the baselines’ tokens do not. | Active |
| LIT-496 | Taming Transformers for High-Resolution Image Synthesis | Esser, Rombach and Ommer (2020), ARXIV-2012.09841. Adds a perceptual loss and a patch-based discriminator to the VQ autoencoder, with an adaptive weight λ = ∇[L_rec] / (∇[L_GAN] + δ) computed from the decoder’s last layer. The result is that f = 16 compression stops destroying perceptual quality, so a 256×256 image becomes 16×16 tokens a transformer can model. Every tokenizer in this line since is this architecture with the codebook retuned. | Active |
| LIT-497 | Autoregressive Model Beats Diffusion: Llama for Scalable Image Generation | Sun et al. (2024), ARXIV-2406.06525. Keeps VQGAN’s encoder-quantizer-decoder and retunes the codebook: ℓ₂-normalized codes, vector dimension 8 instead of 256, size 16384. Utilization goes from 0.29% to 97% and rFID from 9.21 to 2.19. Its other measurement is the one the record needed — the same tokenizer spans rFID 2.19 → 0.70 purely by changing how many tokens the image gets. | Active |
| LIT-498 | MaskGIT: Masked Generative Image Transformer | Chang, Zhang, Jiang, Liu and Freeman (2022), ARXIV-2202.04200. Keeps VQGAN’s tokenizer and replaces its raster-order autoregressive prior with a bidirectional masked transformer decoded in parallel: predict every token, keep the most confident, re-predict the rest. 8 steps instead of 256, up to 64× faster, and better — ImageNet 256×256 FID 6.18 vs 15.78, IS 182.1 vs 78.3. The mask schedule matters and the concave family wins; cosine is best in every experiment. | Active |
| LIT-499 | Neural Discrete Representation Learning | van den Oord, Vinyals and Kavukcuoglu (2017), ARXIV-1711.00937. The paper that made discrete latents trainable: nearest-neighbour lookup into a learned codebook, gradients copied past the non-differentiable step by a straight-through estimator, and two L2 terms — one moving codes toward the encoder, one committing the encoder to its code. ImageNet at 128×128×3 compressed to 32×32 indices over K = 512, a 42.6× bit reduction, with a PixelCNN learned over the indices as the prior. | Active |
| LIT-500 | Vector-quantized Image Modeling with Improved VQGAN | Yu et al. (2021), ARXIV-2110.04627. The origin of the codebook advice SOTA-306 rests on, and it says something sharper than its descendants repeat: factorize the code, so lookup happens in a low-dimensional space while the embedding stays high-dimensional, and ℓ₂-normalize both sides. Its ablation puts codebook usage at 4% at lookup dimension 256, 95% at 16 — and 2% with normalization removed, which is the largest single effect in the table. | Active |
| LIT-501 | The FID Lottery: Quantifying Hidden Randomness in Generative-Model Evaluation | Dufour, Efros and Pérez (2026), ARXIV-2606.20536. Several hundred SiT networks trained on class-conditional ImageNet 256×256, with FID treated as a random variable over a panel of training seeds × sampling seeds. Retraining moves FID 3.2× more than resampling does, the coefficient of variation holds at 1–2% across four model sizes and ten compute budgets, and a lucky training seed reaches the same FID with up to 2× less compute. | Active |
| LIT-553 | Cold Diffusion: Inverting Arbitrary Image Transforms Without Noise | Bansal et al. (2022), ARXIV-2208.09392. Train a network to invert blur, masking, downsampling or snow at every severity, then sample by alternating restoration and re-degradation, and something generative happens without Gaussian noise. The durable result is the sampler, x_{s−1} = x_s − D(x̂₀, s) + D(x̂₀, s−1), which is exact for degradations linear in severity whatever the restorer’s errors. The headline, that noise is not needed, is contradicted by its own Table 5: noiseless blur generation is FID 97.00 on CelebA against 23.11 with noise, and a σ = 0.002 noise sprinkle brings it to 49.45. | Active |
| LIT-554 | Diffusion Forcing: Next-token Prediction Meets Full-Sequence Diffusion | Chen, Marti Monso, Du, Simchowitz, Tedrake, Sitzmann (2024), ARXIV-2407.01392. Train a causal sequence model to denoise tokens that each carry an independent noise level. One model then samples autoregressively, like next-token prediction, and supports guidance over a whole horizon, like full-sequence diffusion. Video rollouts stay stable past 1,000 frames where teacher forcing diverges, shown qualitatively only. On D4RL mazes it beats Diffuser using its own generated actions. A convolutional RNN, not a transformer, at small scale. | Active |
| LIT-555 v2 | Warm Diffusion: Recipe for Blur-Noise Mixture Diffusion Models | Hsueh et al. (2025), ARXIV-2511.16904. Degrade with blur and noise together, and predict the blurry image and the missing high frequencies with two heads. Blur lets the model use the dependency of fine detail on coarse structure. Noise keeps intermediate states on the data manifold. Sweeping the blur-to-noise ratio on CIFAR-10 at 35 steps gives FID 1.97 at BNR = 0 (EDM), 1.85 at BNR = 0.5 and 11.97 at BNR = 10. A small gain, and a large loss as the process goes cold. | Active |
| LIT-559 | Alias-Free Generative Adversarial Networks | Karras et al. (2021), ARXIV-2106.12423. StyleGAN3: fine detail in GAN images sticks to pixel coordinates (“texture sticking”) because the generator aliases. Treat every feature map as a bandlimited continuous signal. Low-pass filter around each nonlinearity at 2× temporary upsampling, feed Fourier features instead of a constant, drop noise inputs. The generator becomes translation-equivariant (EQ-T 63 dB), or also rotation-equivariant (StyleGAN3-R, EQ-R 40 dB), at StyleGAN2’s FID. | Active |
| LIT-560 | Analyzing and Improving the Image Quality of StyleGAN | Karras et al. (2019), ARXIV-1912.04958. StyleGAN2 traces StyleGAN’s droplet artifacts to instance normalization and replaces it with weight demodulation (FID unchanged, artifacts gone). It traces its phase artifacts to progressive growing and replaces growing with a fixed output-skip generator and residual discriminator (FFHQ FID 4.34 → 3.31). It runs R1 only every 16 minibatches at no cost, and adds path-length regularization, which trades FID against smoothness on less structured data. | Active |
| LIT-561 | A Style-Based Generator Architecture for Generative Adversarial Networks | Karras, Laine, Aila (2018), ARXIV-1812.04948. StyleGAN: a mapping network turns the latent z into an intermediate w, and w modulates every synthesis layer through AdaIN. Per-layer noise inputs supply stochastic detail, and mixing two w’s across layers regularizes the model toward scale-specific control. Built on progressive growing. FFHQ FID goes from 8.04 (Progressive GAN) to 4.40, most of it from tuning (5.25) and the mapping network and styles (4.85). Also introduced the FFHQ dataset and the perceptual path length metric. | Active |
| LIT-562 | Projected GANs Converge Faster | Sauer, Chitta, Müller, Geiger (2021), ARXIV-2111.01007. Feed real and generated images through a frozen ImageNet-pretrained EfficientNet, mix its multi-scale features with fixed random 1×1 and 3×3 convolutions, and put a small discriminator on each scale. FID reaches StyleGAN2’s best on LSUN-Church after 1.1M images instead of 88M, and beats it on 22 datasets. A later analysis (LIT-563) shows part of the gain is in FID’s perceptual null space: at equal FID, humans and a CLIP-feature distance prefer StyleGAN2. | Active |
| LIT-563 | The Role of ImageNet Classes in Fréchet Inception Distance | Kynkäänniemi, Karras, Aittala, Aila, Lehtinen (2022), ARXIV-2203.06026. FID’s feature space sits one affine map from ImageNet class logits, so it attends to whatever ImageNet class an image evokes. That is often not the subject: seat belts and bow ties on FFHQ faces. Resampling a fixed generator’s outputs to match those statistics cuts FFHQ FID from 5.30 to 1.78 (−66%) while a CLIP-space distance moves 2.76 → 2.64. A Projected FastGAN with an ImageNet-pretrained discriminator matches StyleGAN2’s FID and loses to it in CLIP space and with humans. | Active |
| LIT-564 | Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks | Zhu, Park, Isola, Efros (2017), ARXIV-1703.10593. Learn G: X → Y and F: Y → X from unpaired collections, with adversarial losses in each domain and a cycle-consistency loss that makes F(G(x)) ≈ x and G(F(y)) ≈ y. The cycle term turns an under-constrained mapping into a usable one for colour and texture changes (horse → zebra, photo ↔ painting, summer ↔ winter). The authors report that it fails on geometric changes. Its own ablation has a one-directional cycle beating the full model on labels → photo. | Active |
| LIT-565 | FLUX-Reason-6M & PRISM-Bench: A Million-Scale Text-to-Image Reasoning Dataset and Comprehensive Benchmark | Fang et al. (2025), ARXIV-2509.09680. 6 million FLUX-generated images with 20 million bilingual captions, organized by six characteristics (imagination, entity, text rendering, style, affection, composition), and with “Generation Chain-of-Thought” breakdowns. Curation cost 15,000 A100 GPU-days. Released with PRISM-Bench, seven tracks scored by vision-language models, used to evaluate 19 text-to-image models. A dataset and a benchmark, with no training ablation read here. | Active |
| LIT-566 | SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis | Podell et al. (2023), ARXIV-2307.01952. SDXL: a 2.6B-parameter UNet latent diffusion model, 3× Stable Diffusion’s, with two text encoders (CLIP ViT-L and OpenCLIP ViT-bigG) and pooled-text conditioning. It is conditioned on each training image’s original size and crop, trained on multiple aspect ratios, and followed by an optional refiner applied SDEdit-style. The one quantitative ablation is size conditioning on class-conditional ImageNet. The headline comparisons are user studies. | Active |
| LIT-572 | FLUX.2: Analyzing and Enhancing the Latent Space of FLUX – Representation Comparison | Black Forest Labs (2025), technical report. DiT-XL flow models on ImageNet 256² in four latent spaces (SD-VAE, FLUX.1-VAE, FLUX.2-VAE, RAE), 30 configurations each: training timestep distribution, training shift and sampling shift, with and without REPA. The training shift moves FID by 61–86%, and without tuning it per space the rankings flip. FLUX.1’s VAE reconstructs far better than SD’s and is less learnable (gFID 10.1 against 7.7). FLUX.2’s VAE, with 8× SD’s dimensionality and semantic regularization, gets both (LPIPS 0.27, gFID 3.7). FLUX.2 is that latent plus Mistral-3 24B and a rectified-flow transformer. | Active |
| LIT-573 | FLUX.1 Kontext: Flow Matching for In-Context Image Generation and Editing in Latent Space | Black Forest Labs (2025), ARXIV-2506.15742. The only paper describing FLUX.1: a 12B rectified-flow transformer in a 16-channel VAE latent, with double-stream then 38 single-stream blocks, fused feed-forward layers and 3D RoPE. Kontext fine-tunes it for editing by appending context-image latents to the sequence, offset by a “virtual time step” in RoPE. The evidence is human preference against other systems and a reconstruction table. It has no ablations. | Active |
| LIT-584 | Accurate structure prediction of biomolecular interactions with AlphaFold 3 | Abramson et al. (Nature 2024), DOI 10.1038/s41586-024-07487-w. AlphaFold 3 predicts complexes of proteins, nucleic acids, ligands, ions and modified residues. A pairformer with lighter MSA processing replaces the Evoformer, and a diffusion module over raw atom coordinates replaces the structure module, with no built-in rotational equivariance. It beats Vina on PoseBusters without the pocket information Vina gets (P = 2.27×10⁻¹³). The generative head hallucinates structure in disordered regions, which cross-distillation from AlphaFold-Multimer reduces. Chirality is violated in 4.4% of PoseBusters predictions even after ranking penalties. | Active |
| LIT-592 | Latent Dirichlet Allocation | Blei, Ng and Jordan (2003), JMLR 3:993–1022. If the words of a document are exchangeable (bag of words), de Finetti’s theorem says the document is a mixture over a latent parameter. LDA makes that parameter a Dirichlet-distributed vector of topic proportions, one per document, with each topic a distribution over words. Fitted by variational EM, it has lower held-out perplexity than a unigram, a mixture-of-unigrams and a pLSI model on two corpora, and the two baselines overfit badly as the number of topics grows. Its 50 topic proportions stand in for 15,818 word features in Reuters classification. | Active |
| LIT-611 v2 | GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium | Heusel et al. (2017), ARXIV-1706.08500. Where FID comes from — and it comes from a subsection of a paper about two time-scale update rules and Nash equilibria. Fit a Gaussian to Inception activations for each distribution and take the Wasserstein-2 distance between them. Validated by monotonicity under six synthetic corruptions. | Active |
| LIT-614 | Improved Techniques for Training GANs | Salimans et al. (2016), ARXIV-1606.03498. Where the Inception Score comes from — exp(E_x KL(p(y|x) ‖ p(y))), proposed as a cheap stand-in for a Mechanical Turk study it correlates with, on 50k samples because part of what it measures is diversity. The incumbent LIT-611 displaced, and the paper that first tied generative evaluation to ImageNet class structure on purpose. | Active |
| LIT-615 | Demystifying MMD GANs | Bińkowski et al. (2018), ARXIV-1801.01401. Where KID comes from, and where the case against comparing FID numbers comes from. The plug-in FID estimator is biased, no unbiased estimator of FID exists, and at d=2048 with 50,000 samples it reliably returns the wrong ordering between two models with a standard deviation small enough to hide it. KID is the squared MMD in the same Inception features, with a cubic kernel and a simple unbiased estimator. | Active |
| LIT-618 v2 | LTX-Video: Realtime Video Latent Diffusion | HaCohen et al., Lightricks (2024), ARXIV-2501.00103. It trades latent size for speed: a 32×32×8 VAE with 128 channels (1:192) moves patchification into the autoencoder, and the decoder does the last denoising step. It generates 5s of 768×512 video in about 2s on an H100 at 20 steps. The denoising decoder, the central idea, is supported by one internal preference study with no reported numbers. | Active |
| LIT-619 v2 | Wan: Open and Advanced Large-Scale Video Generative Models | Wan Team, Alibaba (2025), ARXIV-2503.20314. The open Wan2.1 video models, 1.3B and 14B, and the full recipe behind them. The recipe covers data curation, a 127M causal 3D VAE compressing 4×8×8, a cross-attention DiT trained by rectified flow on an image-then-video resolution curriculum, the 2D context parallelism it needs, and a mask-conditioned image-to-video variant. The abstract claims video scaling laws in data and model size, but the report plots no scaling curve. | Active |
| LIT-620 v2 | HunyuanVideo: A Systematic Framework For Large Video Generative Models | Kong et al., Tencent Hunyuan (2024), ARXIV-2412.03603. A 13B open video model: FLUX-style dual-then-single-stream transformer, a multimodal LLM as text encoder, a from-scratch causal 3D VAE. It fits the line’s only explicit video compute-optimal scaling law, on a proxy model that differs from the shipped one in objective, text encoder and conditioning. It reports no controlled ablations. | Active |
| LIT-621 v2 | Align your Latents: High-Resolution Video Synthesis with Latent Diffusion Models | Blattmann et al., LMU Munich and NVIDIA (2023), ARXIV-2304.08818. Video LDM turns a pretrained image latent diffusion model into a video model by inserting temporal layers and fine-tuning the decoder on video. Starting from the image model halves driving-scene FVD against end-to-end training (534 against 1155). The paper counts 2.2B of the text-to-video model’s 3.1B parameters as trained, and that count leaves out the spatial layers and decoder it also fine-tuned. The abstract describes training only temporal layers. | Active |
| LIT-622 v2 | CogVideoX: Text-to-Video Diffusion Models with An Expert Transformer | Yang, Teng et al., Tsinghua and Zhipu AI (2024), ARXIV-2408.06072. The first widely used open DiT for text-to-video. It has a MAGVIT-style causal 3D VAE, one sequence over text and video with modality-specific adaLN, and full 3D attention. Full 3D attention beats factorized attention at 1.1× to 2.3× the forward time, rising with resolution. The architecture ablations are absent from the first arXiv version and present in the third, and the 2B model ships the position encoding they rejected. | Active |
| LIT-623 v2 | Language Model Beats Diffusion — Tokenizer is Key to Visual Generation | Yu et al., Google and CMU (2023), ARXIV-2310.05737. MAGVIT-v2 is a causal 3D-CNN tokenizer with lookup-free quantization. The causal design lets images and video share one tokenizer, though no experiment uses the video tokenizer for images, and LFQ lets generation keep improving as the vocabulary grows, where VQ turns worse. The title’s “beats diffusion” holds only with guidance, and the “language model” is a masked, MaskGIT-style one. | Active |
| LIT-624 v3 | Step-Video-T2V Technical Report: The Practice, Challenges, and Future of Video Foundation Model | Step-Video Team, StepFun (2025), ARXIV-2502.10248. A 30B text-to-video DiT on a 16×16×8 VAE that reconstructs as well as HunyuanVideo’s 8×8×4 on its own benchmark, though not on Open-Sora 2.0’s, so the model covers 204 frames. It also applies DPO to video (a 55/45 preference score, ties split) and runs at 32% MFU. It reaches parity with Movie Gen and loses to the strongest commercial model it names, despite the abstract’s “state-of-the-art”. | Active |
| LIT-625 v2 | Stable Video Diffusion: Scaling Latent Video Diffusion Models to Large Datasets | Blattmann, Dockhorn, Kulal et al., Stability AI (2023), ARXIV-2311.15127. Video LDM’s architecture is held fixed, and the paper studies what the data does. Curating 4× fewer clips beats the uncurated set, and the advantage survives high-quality fine-tuning. It is the only controlled video-curation study in this line. One of its own filters lost the ablation and was kept anyway. | Active |
| LIT-626 v2 | Movie Gen: A Cast of Media Foundation Models | The Movie Gen team, Meta (2024), ARXIV-2410.13720. A 30B video model built from a Llama 3-style transformer, flow matching and an 8×8×8 temporal autoencoder, with the most controlled ablations in the video line. At 5B, flow matching beats diffusion and video captions beat frame captions. The “LLaMa3 scaling laws” claim rests on image-only runs at four sizes. | Active |
| LIT-627 v2 | Video Diffusion Models | Ho, Salimans et al., Google (2022), ARXIV-2204.03458. The first diffusion model for video: a U-Net factorized into space and time and trained jointly on videos and still frames. Adding eight still frames per video cuts FVD from 202 to 58 at fixed model, batch and steps, though not at fixed frames per step. The abstract’s explanation, lower gradient variance, is never measured. | Active |
| LIT-629 v3 | Self Forcing: Bridging the Train-Test Gap in Autoregressive Video Diffusion | Huang et al., Adobe Research and UT Austin (2025), ARXIV-2506.08009. Self Forcing trains an autoregressive video model on its own rollouts, conditioning on self-generated context and scoring the whole clip with a distribution-level loss. On Wan2.1-1.3B it beats teacher-forced and diffusion-forced training under every objective tried, and holds quality where they degrade frame by frame. It reaches 17 fps at 0.69s latency on one H100. The comparison against slower models is confounded by a 14B teacher. | Active |
| LIT-630 v3 | Flow Matching for Generative Modeling | Lipman et al., Meta AI and Weizmann (2022), ARXIV-2210.02747. A continuous normalizing flow can be trained without simulation by regressing onto a per-sample conditional vector field. Diffusion paths are special cases. The straight “OT” path beats the diffusion path under the same objective, architecture and epochs, and it needs about 60% of the function evaluations. It used uniform timesteps. | Active |
| LIT-631 v3 | From Slow Bidirectional to Fast Autoregressive Video Diffusion Models | Yin et al., MIT and Adobe (2024), ARXIV-2412.07772. CausVid distills a bidirectional video DiT into a 4-step block-causal student with distribution matching, cutting first-frame latency from 219s to 1.3s at 9.4 fps. The controlled result is that a bidirectional teacher makes a better causal student than a causal one. The authors attribute this to error accumulation, but the causal teacher is also the weaker model, and the ablation does not separate the two. The paper does not claim real time, and it reports that the student flickers more and is less diverse. | Active |
| LIT-632 v2 | Make-A-Video: Text-to-Video Generation without Text-Video Data | Singer et al., Meta AI (2022), ARXIV-2209.14792. An unCLIP text-to-image model is extended into video with identity-initialized temporal layers trained on video whose captions are discarded. Text alignment comes entirely from images. That limits it: it “can not learn associations between text and phenomenon that can only be inferred in videos”. The paper has no ablations, and its title describes the training signal, not the data. | Active |
| LIT-633 v2 | AnimateDiff: Animate Your Personalized Text-to-Image Diffusion Models without Specific Tuning | Guo et al., CUHK, Shanghai AI Lab and Stanford (2023), ARXIV-2307.04725. Stable Diffusion 1.5 is frozen and only zero-initialized temporal “motion modules” are trained. The modules then plug into community checkpoints fine-tuned from the same base, with no tuning. Transfer is measured by one table against two baselines not built for the task. The first version’s limitations section, which reports failure on stylized checkpoints, was dropped from the second. | Active |
| LIT-634 v2 | Open-Sora 2.0: Training a Commercial-Level Video Generation Model in $200k | The Open-Sora team, HPC-AI Tech (2025), ARXIV-2503.09642. An 11B video model initialized from FLUX, with most of its compute at 256px and a 768px text/image-to-video stage that is short in iterations but takes 37% of the compute. The “$200k” is one final diffusion-model run at an assumed H200 rental rate. It excludes data, captioning, failed runs, and the FLUX and HunyuanVideo-VAE weights it starts from. A 4×32×32 autoencoder is a further efficiency experiment, and the model adapted to it is unconverged and weaker. | Active |
| LIT-635 v2 | Emu Video: Factorizing Text-to-Video Generation by Explicit Image Conditioning | Girdhar et al., Meta GenAI (2023), ARXIV-2311.10709. Text-to-video is factored into generating an image and then a video conditioned on it, on a frozen text-to-image model. The report runs controlled human-preference ablations of the factorization, zero terminal SNR, low-then-high resolution training and a fine-tune on 1.6K clips filtered on motion, aesthetics and CLIP score together. Movie Gen builds on this recipe. Freezing the image layers beats fine-tuning them only narrowly (55/58). | Active |
| LIT-636 v4 | Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow | Liu, Gong and Liu, UT Austin (2022), ARXIV-2209.03003. It names rectified flow: regress a velocity onto x1 − x0 along the straight interpolation, which beats VP and sub-VP probability-flow ODEs on CIFAR-10 at the same architecture (FID 2.58 against 3.93). The “fast” in the title needs reflow or distillation. Without reflow or distillation one-step FID is 378. Distillation alone gives 6.18. One reflow gives 12.21 undistilled and 4.85 distilled, and full-simulation quality gets worse. | Active |
| LIT-637 v2 | Imagen Video: High Definition Video Generation with Diffusion Models | Ho et al., Google (2022), ARXIV-2210.02303. VDM scaled into a seven-model, 11.6B cascade producing 128 frames at 1280×768. It is the first report to add a separate image-text corpus (60M internal pairs plus LAION-400M) to joint training, and it asserts the benefit without an ablation. Its tested results are that v-prediction beats ε-prediction for video super-resolution, and that distillation to 8 steps is 18× faster at a small metric cost under constant guidance and none under oscillating guidance. | Active |
| LIT-642 | Professor Forcing: A New Algorithm for Training Recurrent Networks | Lamb, Goyal et al., MILA (2016), ARXIV-1610.09038. A discriminator is trained to tell a recurrent network’s hidden-state dynamics under teacher forcing from its dynamics when free-running, and the generator is trained to fool it alongside the usual likelihood loss. Human raters prefer its samples for handwriting generated 20 times past the training length and for raw audio. The likelihood gains are small, the sequential-MNIST claim has no teacher-forcing baseline, and word-level language modeling showed no difference. | Active |
| LIT-643 | One-step Diffusion with Distribution Matching Distillation | Yin et al., MIT and Adobe (2023), ARXIV-2311.18828. DMD distills a diffusion model into a one-step generator by descending an approximate KL(p_fake ‖ p_real), whose gradient is the difference of a frozen “real” score and an online “fake” score. It adds an LPIPS regression loss onto precomputed teacher ODE outputs. It reaches 2.62 FID on ImageNet-64 and 11.49 on zero-shot COCO-30k. The ablation that makes the regression loss look necessary ran its no-regression arm at a different learning rate, and its mode-collapse evidence is qualitative. | Active |
| LIT-644 v2 | Building Normalizing Flows with Stochastic Interpolants | Albergo and Vanden-Eijnden (2022), ARXIV-2209.15571. A continuous normalizing flow can be trained without ODE solves by regressing a velocity onto the time derivative of any interpolant between two densities. It is one of three concurrent origins of this simulation-free objective, with Flow Matching and Rectified Flow. Its path is not the straight line: every experiment uses the trigonometric cos/sin interpolant, which SiT later calls GVP. The image results arrived only in v3, and they are quoted against diffusion baselines, not matched against them. | Active |
| LIT-645 | Stochastic Interpolants: A Unifying Framework for Flows and Diffusions | Albergo, Boffi and Vanden-Eijnden (2023), ARXIV-2303.08797. The earlier interpolant objective is extended with a latent Gaussian term. One learned velocity and score then give an ODE and a family of SDEs, all sharing the interpolant’s marginals, and the SDE’s diffusion coefficient can be chosen after training. For SDEs the regression losses bound the KL, and for the ODE they do not. This is the framework SiT instantiates. The evidence that SDE sampling beats ODE sampling is a 2-D checkerboard and a 128-D Gaussian mixture. The paper reports no image metric and no diffusion baseline. | Active |
| LIT-646 | Improved Distribution Matching Distillation for Fast Image Synthesis | Yin et al., MIT and Adobe (2024), ARXIV-2405.14867. DMD2 drops DMD’s regression loss. It stabilizes training by updating the fake score five times per generator step, adds a GAN loss on real images, and trains multi-step students on their own simulated inputs. The ImageNet ablation is clean: removing the loss costs 2.62 → 3.48 FID, and the update rule restores 2.61. The student is slightly less diverse than its teacher (LPIPS diversity 0.61 against 0.64). The paper does not attribute this to the reverse KL. | Active |
| LIT-647 | Sequence Level Training with Recurrent Neural Networks | Ranzato, Chopra, Auli and Zaremba, Facebook AI Research (2015), ARXIV-1511.06732. The paper that named exposure bias. MIXER anneals a cross-entropy-trained RNN toward REINFORCE on the test metric over its own samples, and beats cross-entropy by 1 to 3 points on three tasks with greedy decoding. Its baselines that avoid exposure bias without a sequence-level reward do not reliably help, so the gain is not shown to come from fixing exposure bias. | Active |
| LIT-648 | Improving Image Generation with Better Captions | Betker et al., OpenAI and Microsoft (2023), no arXiv id. Identical T5-conditioned latent diffusion models are trained on the same images with alt-text, short synthetic or descriptive synthetic captions. The descriptive ones give higher CLIP score, and more synthetic in the blend is better up to the 95% tested. The controlled evidence is CLIP score alone, at 256px, and on ground-truth prompts the gain is inside the checkpoint noise. The paper does not describe the DALL-E 3 model, and it says its DALL-E 3 comparisons are not a measurement of recaptioning. | Active |
| LIT-649 | Scheduled Sampling for Sequence Prediction with Recurrent Neural Networks | Bengio, Vinyals, Jaitly and Shazeer, Google (2015), ARXIV-1506.03099. During training, each previous-token input is replaced by the model’s own prediction with a probability that grows on a schedule. On captioning, parsing and frame-level speech recognition with small LSTMs this beats training on ground truth only, and training only on the model’s own output does badly. The gains are single runs with no variance reported. Feeding uniform random tokens recovers part of the captioning gain, and the paper never measures how errors grow with sequence length. | Active |
| LIT-652 v2 | Video generation models as world simulators | Brooks, Peebles et al., OpenAI (2024), no arXiv id. Sora is a latent diffusion transformer over spacetime patches of a video autoencoder’s latents. It is trained jointly on images and videos at their native durations, resolutions and aspect ratios, with captions from a DALL-E 3-style recaptioner. The report withholds model and implementation details, and it has no table, number, metric or baseline. Every claim, including the scaling comparison and the square-crop ablation, is shown by example videos only. The video reports filed after it cite it and benchmark against it, and it supplies no measurement of its own. | Active |
| LIT-658 | FiT: Flexible Vision Transformer for Diffusion Model | Lu et al. (2024), ARXIV-2402.12376. A DiT variant trained on images without cropping or upsampling. Large images are only resized down to at most 256² pixels at native aspect ratio, and sequences are padded to 256 tokens. Table 3 isolates this pipeline change on DiT-B/2 at 400K steps. At a square 256² output, FID moves 44.83 → 43.34 and sFID worsens 8.49 → 11.11. At 160×320 and 224×448, FID falls from 91.32 and 109.1 to 50.51 and 52.55, but the fixed model never trained on those shapes. | Active |
| LIT-659 | On the Importance of Noise Scheduling for Diffusion Models | Chen, Google Brain (2023), ARXIV-2301.10972. Higher-resolution images are less corrupted by the same per-pixel noise, so the best schedule moves with resolution. It sweeps schedule functions (Table 3) and an input scaling factor b, with x_t = sqrt(γ)·b·x0 + sqrt(1−γ)·ε (Table 4), on ImageNet 64, 128 and 256 with RIN. The best b falls as resolution rises. At 256 with a 1−t schedule, b = 1 gives FID 7.21 and b = 0.4 gives 3.52. This is concurrent with simple diffusion and is the second origin of the resolution-dependent shift. | Active |
| LIT-660 | simple diffusion: End-to-end diffusion for high resolution images | Hoogeboom, Heek and Salimans, Google Brain (2023), ARXIV-2301.11093. Trains single-stage pixel-space diffusion up to 512x512 with four changes. The first is to define the noise schedule at a reference resolution and shift its log-SNR by 2·log(ref/d) at resolution d. Table 2 is the controlled measurement. On ImageNet 256 the unshifted cosine schedule gives FID 7.65 train / 6.87 eval, and the schedule shifted to 32 gives 3.76 / 3.71. This is the origin, with the concurrent Chen (2023), of the resolution-dependent schedule shift SD3 later derives again. | Active |
| LIT-661 | Photorealistic Video Generation with Diffusion Models | Gupta, Yu et al., Stanford and Google (2023), ARXIV-2312.06662. A latent video diffusion transformer over a causal image-and-video encoder. It alternates per-frame spatial attention with local spatiotemporal windows. Two single-run ablations bear on the record’s video practices. Adding a separate corpus of ~970M image-text pairs to ~89M text-video pairs takes zero-shot UCF-101 FVD from 598.8 to 344.5 at 419M (Table 5), with compute not stated. And windowed attention matches full self-attention on UCF-101 (FVD 55.3–59.6 against 59.9) while training 1.7x faster (Table 3b). | Active |
| LIT-676 v4 | DPM-Solver++: Fast Solver for Guided Sampling of Diffusion Probabilistic Models | Lu, Zhou, Bao, Chen, Li and Zhu (2022), ARXIV-2211.01095. The regime SOTA-203 does not mention. At a classifier guidance scale of 8.0 on ImageNet 256×256, FID at 10 function evaluations is 13.04 for first-order DDIM, 114.62 for DPM-Solver-2 and 164.74 for DPM-Solver-3 — higher order is monotonically catastrophic, and the practice’s own framing of DDIM as “the least accurate member of the family” inverts. Two causes, one of which the record already holds as SOTA-202. The fix is order 2, multistep, on the data-prediction parameterization. | Active |
| LIT-677 | Diffusion Models With Learned Adaptive Noise | Sahoo, Gokaslan, De Sa and Kuleshov (NeurIPS 2024), ARXIV-2312.13236. MuLAN replaces VDM’s scalar learned schedule with a per-pixel schedule conditioned on a learned auxiliary latent. At equal steps and architecture the bound moves 2.65 → 2.60 bits/dim on CIFAR-10, and it reaches VDM’s 2.65 in 2M steps instead of 10M. Its own ablations confirm VDM’s invariance where VDM claimed it. A scalar schedule conditioned on the input buys nothing, and a multivariate schedule conditioned on time only is “comparable to that of VDM”. The headline 2.55 is a different estimator. | Active |
| LIT-678 v2 | Diffusion Meets Flow Matching: Two Sides of the Same Coin | Gao, Hoogeboom, Heek, De Bortoli, Murphy and Salimans (2024), a blog post at diffusionflow.github.io. With a Gaussian source, flow matching is a diffusion model with the schedule α_t = 1−t, σ_t = t, and the flow-matching Euler sampler is DDIM. What separates the named recipes is three choices: the loss weighting, the network output and the sampling schedule. “Straight paths” holds only for a single-point target, and a variance-preserving schedule can be straighter for wide data. The content is derivation and toy illustration only: no trained models and no benchmarks. | Active |
| LIT-687 v2 | Fine-Tuning Image-Conditional Diffusion Models is Easier than You Think | Martin Garcia, Knaebel, Schmidt, de Geus, Hermans and Leibe (WACV 2025), ARXIV-2409.11355. Marigold’s “dismal” few-step depth came from the DDIM scheduler’s leading timestep spacing. The spacing hands a pure-noise input a timestep that claims an almost clean one. Switching to trailing spacing makes one step usable (NYUv2 AbsRel 5.7). The fixed single step is still behind the 50-step, 10-member ensemble on all five depth sets. What beats that ensemble everywhere is end-to-end fine-tuning as a one-step model with a task loss, and it works nearly as well from plain Stable Diffusion. | Active |
| LIT-689 | Common Diffusion Noise Schedules and Sample Steps are Flawed | Lin, Liu, Li and Yang (WACV 2024), ARXIV-2305.08891. Common schedules leave signal at the last timestep. Stable Diffusion’s x_T is 0.068·x₀ + 0.998·ε. The model learns to keep the leaked per-channel mean, and pure-noise inference then pins samples to “medium brightness”. Many samplers also never start at t = T. It proposes four fixes: zero terminal SNR, v-prediction, sampling from the last timestep, and a guidance rescale. The only numeric result applies all four at once (COCO FID 22.96 → 21.66). Every single-fix ablation is qualitative, on one to three prompts. It credits the trailing discretization to DPM-Solver, and it never calls the spacing fix “slight”. | Active |
| LIT-691 | Repurposing Diffusion-Based Image Generators for Monocular Depth Estimation | Ke, Obukhov, Huang, Metzger, Daudt and Schindler (CVPR 2024), ARXIV-2312.02145. Marigold fine-tunes Stable Diffusion v2’s U-Net on 74K synthetic images to denoise a depth latent conditioned on the image latent, with the VAE frozen. It samples 50 DDIM steps and ensembles 10 runs, and ranks first on average across five zero-shot depth benchmarks. Its step-count curve, one step at ≈34% NYUv2 AbsRel, was measured under diffusers’ default leading spacing, which the paper never names. That is the scheduler bug LIT-687 corrects. The claim that a generative prior is what helps is not tested. | Active |
| LIT-692 | Understanding Diffusion Objectives as the ELBO with Simple Data Augmentation | Kingma and Gao (NeurIPS 2023), ARXIV-2303.00848. Every standard diffusion loss is ½∫ w(λ) E‖ε̂ − ε‖² dλ for an implied weighting w. Given w, the loss is invariant to the schedule’s interior shape, and the schedule is only an importance sampler. If w falls monotonically in log-SNR, the loss is the expected ELBO of Gaussian-noise-augmented data. App. D.3 derives that flow matching’s weighting equals v-MSE’s under a cosine schedule. In its 64×64 ablations the training schedule makes no difference, and the weighting and the sampler each move FID by 0.1–0.3. Monotone weightings tie EDM. They do not beat it. | Active |
| LIT-693 v3 | Classifier-Free Diffusion Guidance | Ho and Salimans (2022), ARXIV-2207.12598. The technique SOTA-202, SOTA-203 and SOTA-410 all turn on, which this record had never held. Drop the conditioning on a fraction of training examples and one network learns both scores; at sampling, extrapolate away from the unconditional one. The trade it buys is severe and measured: on ImageNet 64×64, w from 0.1 to 4.0 moves FID from 1.55 to 26.22 while IS goes 66.11 to 260.2. Any FID or IS comparison across guidance scales is therefore meaningless, which is the missing source for SOTA-307’s instruction to search the scale per cell. | Active |
| LIT-699 v2 | Diffusion Models Beat GANs on Image Synthesis | Dhariwal and Nichol (2021), ARXIV-2105.05233. The classifier-guidance paper, the ablated U-Net the diffusion literature then treated as the default, and the source of the guidance-scale trade. Its title claim is metric-dependent and the paper says so: ADM-G beats BigGAN-deep on FID and recall at 128, 256 and 512, and loses on precision at all three — 0.78 against 0.86, 0.82 against 0.87, 0.87 against 0.88. It also asserts, once, in the introduction, that the gradient scale can be raised “by an order of magnitude without obtaining adversarial examples”, and runs no experiment anywhere in the paper that tests it. | Active |
| LIT-701 | Discrete Diffusion Modeling by Estimating the Ratios of the Data Distribution | Lou, Meng and Ermon (2023), ARXIV-2310.16834. SEDD: learn the ratios p_t(y)/p_t(x) of a discrete diffusion with score entropy, a Bregman loss that keeps them positive and yields an ELBO. The first discrete diffusion model to reach GPT-2-level zero-shot perplexity, and the absorbing (mask) process beats the uniform one on every table. The GPT-2 comparison is against the released model, trained on different data and re-evaluated without a sliding window. | Active |
| LIT-702 | Simple and Effective Masked Diffusion Language Models | Sahoo et al. (2024), ARXIV-2406.07524. MDLM: the masked-diffusion training objective reduces to a weighted average of BERT-style masked-LM losses, a continuous-time ELBO whose value is invariant to the noise schedule. It is the strongest discrete diffusion language model at its scale — beating SEDD at matched tokens — and still trails a retrained autoregressive baseline in-domain (LM1B 27.04 vs 22.32; OWT 23.21 vs 17.54), which the authors state as the limitation. | Active |
| LIT-703 | Improved Precision and Recall Metric for Assessing Generative Models | Kynkäänniemi, Karras, Laine, Lehtinen and Aila (2019), ARXIV-1904.06991. The definitions SOTA-425 rests on: build a k-NN hypersphere approximation of each distribution’s manifold in a feature space, then precision is the fraction of generated samples inside the real manifold and recall the fraction of real samples inside the generated one. The features are VGG-16 fc2, not Inception — the record had this wrong. Its Figure 4 is the demonstration: two StyleGAN setups at FID 16.9 and 16.7 where one produces good samples at low variety and the other is “practically all images broken”, and the FID-optimised configuration has visible facial distortion. | Active |
| LIT-710 | Efficient Diffusion Training via Min-SNR Weighting Strategy | Hang et al. (ICCV 2023), ARXIV-2303.09556. Weight each timestep’s x0-space loss by min(SNR(t), γ), γ = 5, which is min(γ/SNR, 1) on an ε loss and min(SNR, γ)/(SNR+1) on a v loss. The headline “3.4× faster” is against an x0-prediction baseline with constant weight, on one ViT-B run. Against the ε-prediction default most people train, the UNet gain is 8.55 → 7.32 FID at 200K iterations and 4.21 → 4.14 at 1M. It speeds early training much more than it moves the end point. The “conflicting gradients between timesteps” explanation rests on one fine-tuning probe and a proxy objective. | Active |
| LIT-714 | Analyzing and Improving the Training Dynamics of Diffusion Models | Karras et al. (2023), ARXIV-2312.02696 — EDM2. Two contributions that do not depend on each other. Rebuilding the ADM U-Net so every layer preserves activation, weight and update magnitude takes ImageNet-512 FID from 8.00 to 2.56 at equal compute, the largest single step (6.96 → 3.75) being weight normalization on the learned layers. And post-hoc EMA: store two power-function averages at snapshots and reconstruct any EMA length after training, which shows the optimal length moving with architecture, training time, learning rate, guidance weight and — by a factor of seven under guidance — with which metric you score it. | Active |
| LIT-720 | Analyzing and Improving the Training Dynamics of Diffusion Models | Karras, Aittala, Lehtinen, Hellsten, Aila and Laine (2023), ARXIV-2312.02696. Post-hoc EMA: keep two power-function weight averages during training, snapshot them periodically, and afterwards reconstruct the average for any decay length by least squares — reconstruction error falling as O(1/n⁴) in the snapshot count. That makes the averaging length a sweep instead of a pre-training guess, and the consequence is the headline number here: with the length fixed at 13%, varying the learning-rate decay over t_ref ∈ [30k, 160k] moves FID by up to 72%; sweeping the length post hoc brings the whole bracket within 10% of the optimum. Also EDM2’s magnitude-preserving architecture, reaching FID 1.91 on ImageNet-512 without guidance. | Superseded — by LIT-714; A duplicate. 2312.02696 was already held as LIT-714, filed hours earlier in #375 and merged while this note was being written; LIT-714 is the fuller reading — the whole ablation ladder, the metric-disagreement finding, and NOTE-370 with the appendices — and ten documents cite it against this note’s four, all of which were its own. Retired rather than deleted so the mistake stays visible and the formerly: alias keeps any stray citation resolving. Nothing here is lost: everything this note said, LIT-714 or SOTA-428 says already, including the learning-rate bracket. |
| LIT-721 | Guiding a Diffusion Model with a Bad Version of Itself | Karras, Aittala, Kynkäänniemi, Lehtinen, Aila and Laine (2024), ARXIV-2406.02507. CFG’s quality gain is not the class emphasis. It comes from the unconditional reference model being worse — a harder task on a smaller training slice — so the density ratio p₁/p₀ falls off away from the manifold and guidance points inward. Replace the unconditional model with a smaller, less-trained copy of the same conditional model and the quality gain survives without the diversity loss: ImageNet-512 FID 2.56 → 1.34 on EDM2-S, 1.25 on XXL, 1.01 at ImageNet-64, and unconditional EDM2-S 11.67 → 3.86, which CFG cannot do at all. The mechanism is tested with a negative control: degrade the two models differently and guidance stops helping entirely. | Active |
| LIT-722 | Score-Based Generative Modeling through Stochastic Differential Equations | Song, Sohl-Dickstein, Kingma, Kumar, Ermon and Poole (2020), ARXIV-2011.13456. The framework this record has been standing on without holding. Classifier guidance is Eq. 14 here, six months before ADM: the conditional reverse-time SDE adds ∇ₓ log pₜ(y|x) to the unconditional score, and §5 gives the recipe for the noise-conditioned classifier that supplies it — sample (x(0), y), noise it through the tractable forward SDE, train with a mixture of cross-entropy losses over timesteps. Also the probability flow ODE with exact likelihoods, predictor-corrector sampling, CIFAR-10 IS 9.89 / FID 2.20 / 2.99 bits/dim, and the first 1024×1024 samples from a score model. Its §I.2 imputation approximation is the “replacement” SOTA-397 exists to reject. | Active |
| LIT-743 | Lyra: Generative 3D Scene Reconstruction via Video Diffusion Model Self-Distillation | Bahmani et al. (2025), ARXIV-2509.19296 — Lyra. Train a 3D Gaussian Splatting decoder on the latents of a frozen camera-controlled video diffusion model (GEN3C), supervised only by that model’s own RGB decodings along six camera trajectories per image, with no real multi-view data. Decoding latents rather than pixels is what makes 726 views per scene fit in memory. Single image to 3D beats the reported numbers of ZeroNVS, ViewCrafter, Wonderland and Bolt3D on RealEstate10K, DL3DV and Tanks-and-Temples, and adding real multi-view data to the synthetic supervision does not help. | Active |
| LIT-751 | ”Principal Components” Enable A New Language of Images | Wen, Zhao et al. (2025), ARXIV-2503.08685 — Semanticist. A ViT encoder emits a causal 1D sequence of continuous tokens that conditions a DiT diffusion decoder. During training, every token after a uniformly sampled index is replaced by a learned null token (“nested CFG”), so earlier tokens carry more. The result is coarse-to-fine and PCA-like, and any prefix decodes. ImageNet-256 rFID 0.78 with 256×16 floats against ImageFolder’s 0.80 at 286×32. An autoregressive model trained on only the first 32 tokens reaches gFID 2.57. The diffusion decoder is argued to keep image spectrum out of the tokens, but that is shown against other tokenizers, not ablated. | Active |
| LIT-755 | Wonderland: Navigating 3D Scenes from a Single Image | Liang et al. (2024), ARXIV-2412.12091 — Wonderland (CVPR 2025). Fine-tune CogVideoX-5B-I2V for camera control with a ControlNet branch plus LoRA, then train a reconstructor (LaLRM) that regresses 3D Gaussians directly from the video latents rather than decoded frames. Reading latents beat reading RGB at matched cost (RealEstate10K PSNR 27.10 against 25.06). The reconstructor is trained on captured RealEstate10K, ACID and DL3DV first, then fine-tuned on those plus 20K videos the video model generated, supervised by its own decoded frames; that addition moved benchmark PSNR by at most 0.09. | Active |
| LIT-770 v2 | DMAD: Distribution Matching as Adversarial Distillation for Fast Visual Generation | Yu et al., Texas A&M and ByteDance (2026), ARXIV-2610.02188. DMAD replaces DMD’s fake-score critic with a discriminator and trains the student on a linear loss on its logit. At the unrestricted discriminator optimum that gradient equals DMD’s idealized one (App. A, checked); nothing bounds it off the optimum, where training runs with one discriminator step per student step. One-step ImageNet-64 FID is 1.24 against DMD2’s 1.28, at 570K iterations, and 1.04 with an ImageNet-pretrained discriminator scored only by FID. Four-step SDXL reaches 14.47 FID at lower CLIP. Wan2.1 VBench margins are 0.1–0.3 points. No diversity metric and no seeds are reported. | Active |
| LIT-771 | Glow: Generative Flow with Invertible 1x1 Convolutions | Kingma and Dhariwal, OpenAI (2018), ARXIV-1807.03039. A flow step is actnorm (a per-channel affine layer initialized from one batch), a learned invertible 1×1 convolution in place of Real NVP’s fixed channel permutation, and an affine coupling layer. The 1×1 convolution is the one controlled result: lower CIFAR-10 NLL than reversal or random permutation with either coupling, three seeds, about 7% more wall-clock (Fig. 3). Against Real NVP’s published numbers it is lower on all six datasets (CIFAR-10 3.35 against 3.49 bits/dim), with much larger models. The 256×256 faces are trained on 5-bit images and sampled at temperature 0.7. | Active |
| LIT-772 | Neural Ordinary Differential Equations | Chen, Rubanova, Bettencourt and Duvenaud, Toronto and Vector (2018), ARXIV-1806.07366. Parameterize a hidden state’s derivative with a network, integrate it with a black-box solver, and backpropagate by solving the adjoint ODE backwards, at memory constant in depth (§2, Alg. 1). For densities, the log-density of a state moving under dz/dt = f obeys d log p/dt = −tr(∂f/∂z) (Theorem 1), so a continuous normalizing flow needs a trace, not a determinant. The CNF evidence is 2-D toys. An MNIST ODE-Net matches a ResNet at 0.42% against 0.41% error. | Active |
| LIT-773 | Improving Video Generation with Human Feedback | Liu, Liu, Liang et al., CUHK, Tsinghua and Kuaishou (2025), ARXIV-2501.13918. Three parts. A 182k-pair human preference set over videos from 12 text-to-video systems, rated separately for visual quality, motion and text alignment. VideoReward, a Qwen2-VL-2B reward model trained with a Bradley–Terry loss that models ties. Flow-DPO, Diffusion-DPO rewritten for rectified flow. Carried over exactly, the loss weights the KL by β(1 − t)², which goes to zero at the noise end; that variant raises the reward’s quality scores and drops its text alignment below the base model’s. A constant β fixes it. On an internal 1B video model with LoRA, constant-β Flow-DPO beats SFT and Flow-RWR, and humans prefer it to the base 44% to 27% with 29% ties. Training labels and automatic win rates come from the same reward model, and VBench barely moves. | Active |
| LIT-774 | Diffusion Model Alignment Using Direct Preference Optimization | Wallace, Dang, Rafailov et al., Salesforce AI and Stanford (2023), ARXIV-2311.12908. DPO for diffusion models. The intractable image likelihood is replaced by a bound on the path likelihood, and the reverse chain is approximated by the forward one. The loss compares how much the model beats a frozen reference at denoising the preferred image against how much it beats it on the rejected one, at the same noise and timestep. Trained offline on 851,293 Pick-a-Pic pairs, the tuned SDXL-base is preferred to the untuned base 70.0% of the time on PartiPrompts, and to SDXL base+refiner 69%, by majority vote of five crowdworkers. β is 2000–5000 with the learning rate scaled by 1/β. Fine-tuning on the preferred images alone makes SDXL worse. Checkpoints are chosen by PickScore, which was trained on the same source. There is one run per setting and no intervals. | Active |
| LIT-775 | Inductive Moment Matching | Zhou, Ermon and Song, Luma AI and Stanford (2025), ARXIV-2503.07565. A one-network, from-scratch few-step model that maps the interpolant’s marginal at t to the one at s < t. It is trained by matching, with MMD over groups of samples, the distributions reached from t and from a nearby r. Consistency models are its one-sample, first-moment case, and that case collapses on ImageNet. ImageNet-256 FID is 8.05 at one step and 1.99 at eight, with guidance at sampling time, so 16 evaluations. Training is 1.2M steps at batch 4,096. CIFAR-10 2-step 1.98. One run per row. | Active |
| LIT-776 | Training Diffusion Models with Reinforcement Learning | Black, Janner, Du, Kostrikov and Levine, UC Berkeley and MIT (2023), ARXIV-2305.13301. Treat each denoising step of a stochastic sampler as an action. Each step is then an isotropic Gaussian with an exact likelihood, so policy gradient (DDPO) can fine-tune a diffusion model on any black-box reward of the final image. On Stable Diffusion v1.4, with JPEG compressibility, LAION aesthetic score and a LLaVA + BERTScore alignment reward, both DDPO estimators beat reward-weighted regression per reward query, and the PPO-clipped one is slightly ahead. There is no KL term. Over-optimization is shown and handled by picking checkpoints by hand. Prompt sets are narrow (45–398 animals), there is no human evaluation, and the run count is not stated. | Active |
| LIT-777 | Flow Matching Guide and Code | Lipman et al., FAIR, MIT and Weizmann (2024), ARXIV-2412.06264. A tutorial and PyTorch library, with no new experiments. It states flow matching as “choose a path, regress its generator” across Euclidean, Riemannian, discrete and general Markov-process state spaces. Its claims useful to practice are proved or attributed. The straight conditional path minimizes a bound on kinetic energy and is optimal transport only for a one-point target. Every affine scheduler gives the same exact-ODE endpoint, so a schedule matters only through training and discretization. Sampling t beats weighting by t, citing SD3. It does not cover few-step generation. | Active |
| LIT-778 | DPOK: Reinforcement Learning for Fine-tuning Text-to-Image Diffusion Models | Fan, Watkins, Du, Liu, Ryu, Boutilier, Abbeel, Ghavamzadeh, K. Lee and K. Lee, Google Research, UW–Madison, UC Berkeley, Amazon and KAIST (2023), ARXIV-2305.16381. Policy gradient over the denoising steps, as in DDPO, with a KL to the pre-trained model. The KL on the final image is intractable, so each step’s KL is summed, which bounds it from above. Stable Diffusion v1.5 with LoRA is trained on ImageReward, one model per prompt, against the same reward-weighted objective trained offline on 20K samples. Online RL scores higher on ImageReward and on the held-out aesthetic predictor, and eight raters prefer it. The KL ablation is one prompt. The training reward is also the main metric, and no human compares against the base model. | Active |
| LIT-779 v2 | Flow-GRPO: Training Flow Matching Models via Online RL | Liu, Liu et al., CUHK, Tsinghua and Kuaishou (2025), ARXIV-2505.05470. GRPO on a flow-matching image model. The deterministic sampler is swapped for an SDE with the same marginals, built from the model’s own velocity, which makes each step a Gaussian policy with a closed-form KL. Rollouts use 10 steps and evaluation 40. SD3.5-M with LoRA: GenEval 0.63 → 0.95, OCR accuracy 0.59 → 0.92, PickScore 21.72 → 23.31. Without the KL penalty, quality metrics fall on two tasks and diversity collapses on the third. GenEval is trained on its own templates and scored by its own detector. Quality is judged only by reward models, and there is one run per setting. | Active |
| LIT-780 | FFJORD: Free-form Continuous Dynamics for Scalable Reversible Generative Models | Grathwohl, Chen, Bettencourt, Sutskever and Duvenaud, Toronto, Vector and OpenAI (2018), ARXIV-1810.01367. Replaces the CNF’s exact Jacobian trace with Hutchinson’s estimator, one noise vector fixed per solve, so the log-density costs O(D) per evaluation and the dynamics network is unrestricted (§3, Eq. 8). Best reversible model on all five tabular sets, behind the autoregressive MAF-DDSF on four. MNIST 0.99 bits/dim against Glow’s 1.05; CIFAR-10 3.40 against Glow’s 3.35, with under 2% of Glow’s parameters. Training took about five days on six GPUs, and NFE grows during training until it can become “prohibitive” (§6). | Active |
| LIT-781 | Density estimation using Real NVP | Dinh, Sohl-Dickstein and Bengio, Google Brain (2016), ARXIV-1605.08803. The affine coupling layer: half the dimensions pass through unchanged and parameterize a scale and shift of the other half, so the Jacobian is triangular, its log-determinant is a sum, and the inverse costs the same as the forward pass whatever networks compute the scale and shift (§3.2–3.3). Stacked with squeeze and factor-out in a multi-scale architecture, it gives exact likelihood, exact latent inference and one-pass sampling. It does not beat PixelRNN’s likelihood on any dataset it shares with it (CIFAR-10 3.49 against 3.00 bits/dim, Table 1), and sample quality is shown, not measured. | Active |
| LIT-782 | Pick-a-Pic: An Open Dataset of User Preferences for Text-to-Image Generation | Kirstain, Polyak, Singer, Matiana, Penna and Levy, Tel Aviv University and Stability AI (2023), ARXIV-2305.01569. A web app that generates two images per prompt and logs which one the user keeps, or a tie. The paper’s version holds 583,747 training pairs over 37,523 prompts from 4,375 users. PickScore is CLIP-H fine-tuned on those pairs with a reward-model loss. It predicts held-out user choices at 70.5% under a tie-aware accuracy where chance is 56.8%, against 68.0% for the authors’ expert annotators. Its ranking of 45 model configurations correlates 0.790 with real users’ Elo. On MS-COCO captions FID ranks nine configurations at −0.900 against human experts, because people prefer higher guidance. Every validation is in-distribution or uses colleague annotators. Diffusion-DPO trains on this data and Flow-GRPO optimizes PickScore as a reward. | Active |
| LIT-783 | Large Scale Diffusion Distillation via Score-Regularized Continuous-Time Consistency | Zheng et al., Tsinghua and NVIDIA (2025), ARXIV-2510.08431. sCM made to run on 14B text-to-image and text-to-video models, with a FlashAttention-2 JVP kernel under FSDP and context parallelism. Pure sCM loses fine detail and temporal coherence there. Adding a DMD loss at weight 0.01 (rCM) fixes it. Four-step Wan2.1 VBench is 84.43 at 1.3B against a re-implemented DMD2’s 84.56, and 84.92 at 14B; GenEval tracks the teacher at 1–4 steps. The diversity advantage over DMD2 is shown in samples only. One run per row. | Active |
| LIT-784 | Mean Flows for One-step Generative Modeling | Geng, Deng, Bai, Kolter and He, CMU and MIT (2025), ARXIV-2505.13447. Train a network for the average velocity u(z, r, t) over an interval, using the identity u = v − (t − r)·du/dt with du/dt from a Jacobian-vector product and the target under stop-gradient. With r = t it is flow matching. From scratch on ImageNet-256 with DiT-XL/2 and guidance folded into the target, one-step FID is 3.43 at 240 epochs and two-step 2.20 at 1,000. The best one-step baselines it beats, Shortcut 10.60 and IMM 7.77, are copied rows at other budgets. Behind iCT on CIFAR-10. One run per row. | Active |
| LIT-785 | STARFlow: Scaling Latent Normalizing Flows for High-resolution Image Synthesis | Gu et al., Apple (2025), ARXIV-2506.06276. TarFlow moved into a VAE latent with a deep-shallow split: one deep Transformer flow block and five 2-layer ones. It adds a decoder fine-tuned on noisy latents in place of score denoising, and a guidance rule derived from the Gaussian score. ImageNet 256 FID 2.40 at 1.4B against a retrained pixel TarFlow’s 5.56, with SiT’s quoted 2.06 and MAR’s 1.55 still ahead. ImageNet 512 3.00 against EDM2’s 1.25; COCO zero-shot 9.1; GenEval 0.56. The only head-to-head against diffusion is a plot against an in-house 2.1B DiT trained on 200M images (STARFlow’s schedule is 400M) with a stock decoder. Sampling is about 2.2 s per 256×256 image with guidance. | Active |
| LIT-786 | Improved Techniques for Training Consistency Models | Song and Dhariwal, OpenAI (2023), ARXIV-2310.14189. Consistency training from random weights, without a teacher or LPIPS. The teacher network takes no EMA, because with one the limiting objective carries no information about the data. A Pseudo-Huber loss replaces LPIPS, whose ImageNet features can leak into FID. The step count doubles on a schedule, and noise levels are drawn log-normally. One-step FID is 2.51 on CIFAR-10 and 3.25 on ImageNet-64, two-step 2.24 and 2.77, beating consistency distillation. Ablations are CIFAR-10 curves, one run each. | Active |
| LIT-787 | One Step Diffusion via Shortcut Models | Frans, Hafner, Levine and Abbeel, UC Berkeley (2024), ARXIV-2410.12557. A flow-matching network also conditioned on step size d, trained in one run: the flow-matching loss at d = 0, and a target for step 2d built from two of its own d-steps. On DiT-B at matched compute, one model gives 6.9 / 13.8 / 20.5 FID at 128 / 4 / 1 steps on CelebA-HQ-256 and 15.5 / 28.3 / 40.3 on ImageNet-256. Progressive distillation is better at one step (14.8 and 35.6) and cannot sample at many. The re-run consistency baselines are far below their published numbers. DiT-XL reaches 10.6 in one step. One run per cell. | Active |
| LIT-788 | ImageReward: Learning and Evaluating Human Preferences for Text-to-Image Generation | Xu, Liu, Wu, Tong, Li, Ding, Tang and Dong, Tsinghua, Zhipu AI and BUPT (2023), ARXIV-2304.05977. A reward model for text-to-image preference: a BLIP backbone trained on 136,892 ranked pairs over 8,878 DiffusionDB prompts. It picks the rater’s choice 65.14% of the time, against 54.82% for CLIP similarity, about as often as one annotator agrees with another (65.3%). ReFL, added in v3, backpropagates the reward through one denoising step taken late in the chain. On Stable Diffusion v1.4 it is ranked over the base model 58.79% of the time, ahead of filtering, reward weighting and RAFT. Best-of-64 under the same reward reaches 73.33%. Its premise, that a diffusion model has no likelihood for RLHF to use, is true of the whole chain and not of a step. | Active |
| LIT-789 | Multisample Flow Matching: Straightening Flows with Minibatch Couplings | Pooladian et al., Meta AI, NYU and Weizmann (2023), ARXIV-2304.14772. Flow matching with noise and data paired within each minibatch by OT (BatchOT) or a cheaper stable matching keeps the marginals exact. As batch size grows the paths straighten toward the OT map. On ImageNet-64 the FID-20 point drops from 29 Euler steps to 12, and adaptive FID from 13.93 to 12.37, at equal likelihood and 4% more time per iteration. Low-NFE numbers stay far from one-step quality: 38.86 FID at 4 Euler steps on ImageNet-32. One run per row. | Active |
| LIT-790 | Simplifying, Stabilizing and Scaling Continuous-Time Consistency Models | Lu and Song, OpenAI (2024), ARXIV-2410.11081. Continuous-time consistency models had been unstable. A trigonometric flow parameterization, identity time input with positional embeddings, normalized tangents, learned loss weights and a Flash Attention JVP make them train to 1.5B parameters. Continuous time beats every discrete step count tried. Two-step FID is 2.06 on CIFAR-10, 1.48 on ImageNet-64 and 1.88 on ImageNet-512, against teachers at 2.01, 1.33 and 1.73. Every model, “consistency training” included, starts from the pretrained diffusion model. | Active |
| LIT-791 | Improving the Training of Rectified Flows | Lee, Lin and Fanti, CMU and Microsoft Research (2024), ARXIV-2405.20320. One reflow round is enough, argued informally, and the 2-rectified flow then trains better with a U-shaped timestep distribution and an LPIPS-Huber loss. Started from EDM, with no separate distillation stage, one-step FID is 3.07 on CIFAR-10 and 4.31 on ImageNet-64. Its own table lists DMD at 2.62 on ImageNet-64. There is no distillation-alone arm from the same teacher, and many-step quality is not compared with the pre-reflow model. | Active |
| LIT-792 | Discrete Flow Matching | Gat et al., FAIR and Weizmann (2024), ARXIV-2407.15595. Flow matching on discrete tokens as a continuous-time Markov chain, with velocities of the same form as the continuous case, any coupling, a family of schedulers, and corrector sampling that can re-mask emitted tokens. At 150M its generative perplexity beats SEDD’s in float32 (GPT-2 29.3 against 40.5 at 1,024 NFE). In double precision every discrete model’s perplexity rises 2–3.5 times and the gap narrows (97.5 against 109.5). At 1.7B on code it reaches HumanEval pass@1 6.7 against a same-data autoregressive model’s 14.3. | Active |
| LIT-793 | Flow map matching with stochastic interpolants: A mathematical framework for consistency models | Boffi, Albergo and Vanden-Eijnden, Courant Institute (2024), ARXIV-2406.07507. Learn the two-time flow map X_{s,t} of a probability-flow ODE. Distillation from a velocity field can use a Lagrangian loss (new) or an Eulerian one, which is the continuous-time limit of consistency distillation; both bound a Wasserstein distance. Consistency training’s update is shown not to descend any loss. On CIFAR-10 the Lagrangian loss distils far better than the Eulerian: FID 7.13 against 48.32 at two steps, teacher 5.53. Direct training on ImageNet-32 gives 16.90 FID at four steps. Small models; no one-step image table. | Active |
| LIT-794 | Normalizing Flows are Capable Generative Models | Zhai et al., Apple (2024), ARXIV-2412.06329. TarFlow stacks causal Transformers over image patches as affine autoregressive flows, reversing the order between blocks, trained by exact maximum likelihood. A likelihood model reaches 2.99 bits/dim on unconditional ImageNet 64×64, below every diffusion and autoregressive number it quotes (Table 2). Sampling needs three additions: Gaussian input noise about 25× the dequantization noise, a Tweedie denoising step using the flow’s own score, and classifier-free guidance on the affine parameters. Best FIDs are 2.66 at ImageNet 64 and 5.03 at 128, against quoted 1.55 (EDM) and 1.94. Sampling is sequential: 32 samples at 64×64 take about two minutes on one A100. | Active |
| LIT-796 | InstaFlow: One Step is Enough for High-Quality Diffusion-Based Text-to-Image Generation | Liu, Zhang, Ma, Peng and Liu, UT Austin and Helixon (2023), ARXIV-2309.06380. Reflow Stable Diffusion once, then distil the reflowed model to one step with an LPIPS regression. Against distilling SD 1.4 directly, at the same 100K steps and 3.2M generated pairs, reflow first gives one-step FID 31.0 against 40.9 on COCO-2017-5k and 20.0 against 34.6 on COCO-2014-30k. The scaled model, InstaFlow-0.9B, scores 23.4 and 13.10. The teacher is a diffusion model, not a rectified flow, and the distiller is a regression, not distribution matching. One run per row. | Active |
| LIT-797 | Aligning Text-to-Image Models using Human Feedback | Lee, Liu, Ryu, Watkins, Du, Boutilier, Abbeel, Ghavamzadeh and Gu, Google Research and UC Berkeley (2023), ARXIV-2302.12192. RLHF’s recipe carried to a text-to-image model, without the RL. Two labellers mark 27,528 Stable Diffusion v1.5 images good or bad on templated count, colour and background prompts. A small head on frozen CLIP embeddings learns that label. One offline round of reward-weighted likelihood plus a pre-training loss then fine-tunes the model. Alignment wins 50% of 120 prompts by a two-thirds vote against 3% for the base. Fidelity loses: FID 13.97 to 16.76, and 3% against 20% when compared with best-of-16 sampling from the base under the same reward. One run, no diversity measure. | Active |
| LIT-798 | Improving and generalizing flow-based generative models with minibatch optimal transport | Tong et al., Mila (2023), ARXIV-2302.00482. Flow matching trains on any coupling of source and target samples. Pairing each minibatch by exact optimal transport (OT-CFM) gives near-OT paths in 2-D (normalized path energy 0.02–0.09 against 0.2–2.7 for the independent coupling, five seeds) at under 1% overhead. On CIFAR-10 its gain is small: adaptive-solver FID 3.58 against 3.66 at 134 against 146 NFE, and the independent coupling is better at 1,000 Euler steps. Its rerun of the path comparison, at a recipe that beats Flow Matching’s own, has the straight path ahead of VP at 100 Euler steps (4.64 against 7.77). | Active |