Key takeaways
- The premise is that rectified flow had better theory and no mandate. Straight-line paths between data and noise had been shown to help in small and medium class-conditional experiments — LIT-447 is the one the paper names — and had not been established at scale or for text-to-image.
- The systematic study is the load-bearing part: 61 formulations — epsilon and v prediction, several noise schedules, rectified flow with several timestep distributions — swept at matched settings and ranked. Rectified flow with a logit-normal timestep distribution wins; rectified flow with uniform timesteps does not, which is the distinction that matters and the one a summary would drop.
- Rectified flows are sample-efficient, clearly better at few sampling
steps; by 25 steps only the best-tuned variant stays ahead of
eps/linear. So the advantage is largest exactly where inference cost binds. - Resolution-dependent timestep shifting. More pixels need more noise to destroy the same amount of signal, so a timestep at one resolution must be mapped to a different timestep at another to corrupt equivalently. Without the shift, a schedule tuned at low resolution is wrong at high resolution in a predictable direction.
- MMDiT: separate weights for the text and image streams — separate projections and MLPs per modality — with attention run over the concatenated sequence so information flows both ways. Beats UViT and DiT on the same budget, and improves typography, prompt following and human preference specifically.
- Predictable scaling, and validation loss correlates with human preference and with the automated text-to-image metrics — which is what licenses choosing a configuration by loss before spending on the large run.
- QK-RMSNorm to survive high resolution. Mixed-precision training diverged when moving to high resolution; normalizing Q and K before attention, following the ViT literature’s attention-entropy diagnosis, fixed it.
- More latent channels in the autoencoder (16) improve reconstruction and scale better, because the harder prediction task rewards capacity.
- Largest model 8B, weights and code released.
Standing in the anthology
It closes a gap the record had left open while naming it. SOTA-188 already lists Stable Diffusion 3 as an implementation, and the record did not hold the paper. The listing was right: SD3’s logit-normal over timesteps is SOTA-188’s “concentrate the training noise in the middle” in rectified-flow coordinates, and this is now an 8B confirmation of that half of the practice in a formulation EDM did not test.
Two instructions here were new to the record. Resolution-dependent timestep shifting (SOTA-263) was not new to the field. Simple diffusion (LIT-660) and Chen (LIT-659) derived and measured it 14 months earlier, and §5.3.2 here says the shift is “similar to (Hoogeboom et al., 2023)”. What this paper adds is the rectified-flow derivation and a human-preference sweep of the sampling shift (Fig. 6), which chose α = 3.0 where its own formula gives 4. And MMDiT (SOTA-262) is the record’s first document on how a multimodal generative backbone should divide its parameters.
What it builds on. The straight path is LIT-636’s. SD3’s rectified flow is a straight line between data and standard normal noise, with the network regressing the velocity, credited jointly to Liu et al., Albergo and Vanden-Eijnden, and Lipman et al. Its uniform-timestep variant, “used widely in previous works on Rectified Flows”, is LIT-636’s own setting. What this paper changes is where on that path training is spent: the logit-normal, mode and CosMap timestep densities. The 61-way study then shows that the uniform version does not win. The study’s common frame is LIT-692’s. Every formulation is written as a weighted loss in log-SNR “following Kingma & Gao (2023)”, and the implied weightings for EDM and for v-prediction under the cosine schedule are taken from that paper. That frame lets a new timestep density be recast as a weighting (eq. 18), so that all 61 formulations are ranked on one axis.
MMDiT was later tested directly. CogVideoX (LIT-622) compared its expert adaLN, where text and video share one sequence and differ only in their adaLN, against MMDiT, evaluated on 500 WebVid test videos. It ran one MMDiT arm at equal parameters and one at equal depth with twice the parameters. It reports that expert adaLN “significantly outperforms” MMDiT at equal parameters on FVD, CLIP4Clip and loss, with no values given. It skipped a cross-attention DiT, citing this paper for that arm being inferior to MMDiT.
The QK-normalization finding is a rediscovery, and worth recording as one. The record already holds QK-norm from the language-model side (SOTA-131, SOTA-192, SOTA-050). Here it arrives independently, from a divergence at high resolution in mixed precision, diagnosed through the ViT literature rather than the LLM one. That two lines reached the same fix from different failures is the sort of thing DP-007 says has no author and so no arrival event — except that here there are two, and neither cites the other’s.
What it does not settle is whether the straight-line path is better because it is straight. The study ranks formulations and does not isolate a mechanism; LIT-447’s transport-cost argument is the nearest thing to one and it is an observation about path length, not a proof.
Two later papers measure against this line. Flow-GRPO (LIT-779) takes SD3.5-Medium as its base model and trains it with online RL. GenEval goes from 0.63 to 0.95, on prompts from GenEval’s own templates scored by its own detector, and the published SD3 rows anchor its Table 1. STARFlow (LIT-785) quotes SD3’s GenEval of 0.74 against its own 0.56, the widest of its text-to-image gaps.