Black Forest Labs (2025) — url
Key takeaways
- The trade-off: a latent space is judged on learnability (how well a generator learns in it), quality (reconstruction) and compression. They conflict
- The setup: a DiT-XL flow model on ImageNet 256², with a constant learning rate of 1e-4, batch 256, 50 Euler steps and FID-50k. Each autoencoder gets 3 timestep distributions (shifted uniform, logit-normal, plateau logit-normal) × 5 training shifts × 5 sampling shifts, with and without REPA. Sequence length is held at 256 tokens
- Reconstruction (Table 1, LPIPS / PSNR / rFID): RAE 1.67 / 18.8 / 0.61. SD 0.95 / 25.1 / 0.65. FLUX.1 0.34 / 31.1 / 0.18. FLUX.2 0.27 / 31.5 / 0.11. RAE’s rFID is good while its reference-based metrics are the worst, which matters for editing
- Learnability (Figure 1, gFID at the best configuration): RAE 3.10, FLUX.2 3.70, SD 7.73, FLUX.1 10.13. Widening FLUX.1’s bottleneck (4× the latent dimension, weaker KL) for editing fidelity cost learnability
- Tuning matters more than the representation, at the margin: the best-to-worst training-shift gap is 61.5–86.4% relative FID. The training distribution gap is 6.8–36.8%, with logit-normal or plateau logit-normal always beating shifted uniform. The sampling-shift gap is 4.5–38.7%. “Without shifting, RAE performs worse than the FLUX.2 AE with shifting, whereas RAE outperforms the FLUX.2 AE if both are evaluated with optimal parameters”
- REPA improves every latent space, RAE included
- FLUX.2: Mistral-3 24B VLM and a rectified-flow transformer (SwiGLU, a global modulation mechanism), in this latent. Human-preference win rates against open-weight models are 59.8–66.6%
Standing in the anthology
Filed from #163 (the stable diffusion lineage, “flux2”). FLUX.2 has
no paper. This report is BFL’s technical account of it, and the
announcement post adds nothing measurable. It extends Kontext
(LIT-573), whose VAE it re-examines. It sources SOTA-346.
The √(m/n) claim is weaker than the report makes it sound. It says the optimal training shift for SD and RAE is “consistent with” α = √(m/n) from latent dimensionality. But SD’s optimum is α = 1, the baseline itself. RAE’s predicted √(768/16) = 6.93 is the largest value in the grid, and the report says RAE’s true optimum “might be even higher”. FLUX.2 prefers 4.63 against a prediction of 2.82. This bears on SOTA-263, whose correspondence it extends from resolution to channels without testing it.
Read — NOTE-311.