Summary

Final cleanup PR for the inpainting feature merged in #2309 (#2319, #2320 covered earlier follow-ups). Bundles six small, independent items:

• sdxl_train.py — drop leftover .reshape(batch[\"images\"].shape) no-op in the masked-image VAE encode path. Dead since the ff7945d batched-mask fix.
• model_util.expand_unet_to_inpainting — replace the isinstance(unet.config, dict) guard with unet.register_to_config(in_channels=9) so the diffusers FrozenDict actually gets updated. No functional impact (downstream reads unet.in_channels), but the config attribute is now consistent.
• sample_image_inference — branch the resolution rounding by pipeline class: SDXL has 2 downsamples (latent /4 → image /32), SD1.5/2.x has 3 (latent /8 → image /64). Also fixes the inaccurate "SDXL requires latents divisible by 8" comment.
• inpainting_minimal_inference.py — remove the unused denoise() function (actual loop is inlined in main()).
• BaseDataset.random_mask — drop unused ratio / mask_full_image parameters and the now-dead cloud_mask import; update the single caller.
• --img → --i — rename the per-prompt directive for consistency with Musubi Tuner, which already uses --i alongside the shared w/h/l/d/s directives. Updates docs (EN/JA), the parser, the warning string, and the comment.
  • Bonus fix: when --train_inpainting is set but the prompt has no --i, the code previously warned and fell through to the standard pipeline, which crashes the 9-channel UNet on 4-channel input. Now warns and skips the sample, matching the existing missing-file skip behaviour.

Items #10 and #11 from the original review notes were dropped: #10's described problem doesn't actually exist (the script already splits vae_dtype/weight_dtype correctly), and the gen_img.py --i precedent referenced in the original #11 came from a sibling repo (Musubi Tuner) — the rename is still useful for cross-repo consistency.

Test plan

• # 4 — sdxl_train inpainting training step runs
• # 6 — training with diffusers UNet path runs (config update no longer raises)
• # 7' — verified resolution gets rounded as expected
• # 8 — inpainting_minimal_inference.py still runs end-to-end
• # 9 — training with --train_inpainting runs (caller path exercised)
• # 11 — --i works; --img is silently ignored; missing --i now warns + skips instead of crashing the 9ch UNet

🤖 Generated with Claude Code