Activity

1. tutao0123 commented on Sep 14, 2026

   

   tutao0123

   on Sep 14, 2026

   Contributor

   More actions

   This pattern often comes from train/infer mismatch on how the source–target pair is resized and conditioned, not from LoRA rank alone.

   Worth checking in order:

   1. Pair geometry — are pre/post images exactly the same H×W before encode? Letterbox vs center-crop vs stretch will teach the model a soft “border expand”.
   2. Encode path — if source and target go through VAE with different pad/crop, latent grids won’t share the same pixel lattice and you get the slight structural drift you describe.
   3. Noise / CFG — edit LoRAs are sensitive to train noise scale and to high CFG at inference; try identity or near-identity prompts with lower CFG and see if the border still grows.
   4. Dataset edge padding — constant-color pads around the canvas are a classic outpainting prior.

   A quick A/B: same-size center crop (no pad), train a short run, then infer with identical resolution + low CFG. If the halo shrinks, it’s preprocessing; if not, share your train YAML (resolution, bucket, noise, CFG) and one sample pair sizes and we can narrow further.
2. lllllialois commented on Oct 14, 2026

   

   lllllialois

   on Oct 14, 2026

   Author

   More actions

   Thanks for the suggestions! The border expansion issue has been resolved after adjusting the preprocessing (ensuring consistent geometry). However, pixel-level alignment remains a challenge.

   I have verified that the training and inference settings are identical, but I’ve observed a significant discrepancy between frameworks using the official flux.2-klein-base-9b model:

   With the same prompt and identical parameters, DiffSynth results show much less deviation from the source image, whereas Diffusers produces significantly larger changes.