FLUX LoRA Fine-Tuning: Border Outpainting and Pixel-Level Misalignment Issues in Image Editing

[lllllialois]

Why does the flux.2-klein-base-9b model, when fine-tuned via LoRA on pixel-aligned pre- and post-edited image pairs with prompts for an image editing task, exhibit slight border expansion (outpainting), a failure to preserve pixel-level structural consistency, and minor detail distortions during inference?

[tutao0123]

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.