Black Forest Labs (2025) — ARXIV-2506.15742
Key takeaways
- FLUX.1 (§2): a rectified-flow transformer in the SD3 lineage (LIT-449). The convolutional autoencoder is trained from scratch at 16 latent channels. The transformer has double-stream blocks (separate weights for text and image, joint attention), then 38 single-stream blocks. The single-stream feed-forward is fused with the attention projections for larger matmuls. 3D RoPE indexes (t, h, w)
- Reconstruction (Table 1, 4096 ImageNet images): FLUX-VAE PSNR 31.1, against 29.6 for SD3’s VAE and 25.9 for SDXL’s
- Kontext (§3): context-image tokens are appended to the target tokens, with RoPE time index i for the i-th context image. It is trained with the rectified-flow loss from a text-to-image checkpoint on “millions” of curated pairs, then distilled with LADD for few-step sampling. Channel-wise concatenation “was also tested” and did worse in initial experiments. No numbers are given
- Appendix A.2: SD3’s resolution shift α of timesteps sampled from a logit-normal is the same as a logit-normal with μ + log α. This is the identity the FLUX.2 report (LIT-572) builds its sweep on
- Evaluation: KontextBench, 1026 crowd-sourced image-prompt pairs, with human preference and latency against GPT-Image-1, Gen-4 and others. It leads on local and text editing and character reference, and is second on global editing and style reference
Standing in the anthology
Filed from #163 (the stable diffusion lineage, “flux1”). FLUX.1
itself was released without a paper, and this is the document BFL asks to be
cited for it. It extends SD3 (LIT-449) and is extended_by the FLUX.2
latent-space report.
No practice. Every design choice here is asserted, not ablated, and the comparisons are against other systems. Sequence concatenation for editing is widely adopted since, and that adoption is not evidence (DP-005).
Read — NOTE-310.