The authors studied CFG in bespoke TI2I flow matching models trained specifically for editing, rather than adapted from T2I generators.
Correcting classifier-free guidance in early steps improves image editing fidelity
Authors show that CFG's unconditional branch approximates mean edits instead of no-edit, and replacing it for early flow steps enhances the alignment-preservation tradeoff without extra cost.
Big Tech
Andreas Floros · Elena Koritska · Teofilo de Campos · Christoph Vogel · Chu-Tak Li · Anuj Sharma · +1 more
Imperial College London · Microsoft
Research Digest··2 min read
Floros et al.
Why this paper
From Microsoft and Imperial College London
In one line
Standard classifier-free guidance harms source preservation in early, noisy flow steps for text-and-image-to-image editing; replacing the unconditional velocity with a no-edit velocity fixes it.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors (2 noted)
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§