Migration Notes (SDXL -> Flux)
Reused As-Is
- Trainer loop structure
- Accelerator stack and checkpoint flow
- Pairwise criterion math and evaluation logic
- Dataset filtering and pseudo-preference pipeline
Flux-Specific Changes
- Base model switched to FLUX.1-schnell.
- Diffusion UNet path replaced with Flux transformer token path.
- Latent handling changed to packed latents + latent image ids.
- Second tokenizer switched from CLIP tokenizer to T5 tokenizer.
Config Changes
- step_flux_base now points to Flux checkpoints.
- guidance_scale default set to 0.0.
- Removed SDXL-only model fields from Flux run config.
Smoke Test Policy
- Use /home/user/aev/bin/python for checks.
- Start with syntax/import smoke tests before full training launch.