Contribution
The only published description of FLUX.1’s architecture. Kontext shows that one flow model can do text-to-image and in-context editing by treating context images as extra tokens.
Key insight
Editing as in-context generation. A context image in the same latent space is appended to the sequence, with a RoPE time offset. The model can then attend to it at any resolution or aspect ratio, and more than one image fits without architectural change.
Key results
- FLUX-VAE reconstruction: PSNR 31.1 and LPIPS 0.332, against SD3’s VAE at 29.6 and 0.452 (Table 1)
- Human preference on KontextBench: best on local editing, text editing and character reference, and second on global editing and style reference. Fastest latency among the compared APIs (Figure 7)
- The logit-normal shift identity: μ = log α (appendix A.2)
Limitations
- No ablations. The one design comparison, channel-wise concatenation, is a sentence with no numbers
- Human preference against black-box systems on the authors’ own benchmark
- Multi-turn editing still degrades, as §5 says