* Fix ControlNetDataset delegate: forward train_inpainting to DreamBoothDataset

ControlNetDataset.__init__ constructs an internal DreamBoothDataset via
positional arguments. PR #2309 inserted a new train_inpainting parameter
into DreamBoothDataset.__init__ between prior_loss_weight and
debug_dataset, but this delegate call was not updated, so every
subsequent positional argument was shifted by one slot:

  - train_inpainting       <- received debug_dataset (bool, type-correct
                              but logically wrong: enables inpainting whenever
                              debug mode is on)
  - debug_dataset          <- received validation_split (float)
  - validation_split       <- received validation_seed (int)
  - validation_seed        <- received resize_interpolation (str)
  - resize_interpolation   <- received skip_image_resolution (tuple)
  - skip_image_resolution  <- defaulted to None (missing)

This breaks ControlNet training entirely. Forward the dataset's own
train_inpainting attribute to keep the delegate consistent.

* Fix train_textual_inversion mask interpolate to match new batch shape

Commit ff7945d (PR #2309) fixed the masks/masked_images batch shapes from
[B,1,1,H,W]/[B,1,C,H,W] to [B,1,H,W]/[B,C,H,W] and replaced the per-item
F.interpolate loop with a single batched call in fine_tune.py, train_db.py,
train_network.py, and sdxl_train.py — but train_textual_inversion.py was
missed.

After ff7945d batch["masks"] is [B,1,H,W]; iterating over it yields
[1,H,W] slices. F.interpolate on a 3D tensor with a 2-element size argument
performs 1D interpolation and errors on the dimension mismatch, so
textual-inversion inpainting training crashes on the first step.

Replace the loop+stack with the same single-call form used by the other
training scripts, and drop the now-redundant .reshape() workaround.

* doc: avoid typo warning in documentation for mask visualization output

Fix masked-image VAE encode dtype in fine_tune and train_textual_inversion

- sdxl_train: drop leftover .reshape(batch["images"].shape) no-op in
  masked-image VAE encode (dead since the ff7945d batched-mask fix).
- model_util.expand_unet_to_inpainting: use unet.register_to_config(in_channels=9)
  so the diffusers FrozenDict stays in sync (the previous isinstance(..., dict)
  guard either silently skipped or raised, depending on diffusers version).
- sample_image_inference: round resolution to mod-32 for SDXL (2 downsamples,
  latent /4) and mod-64 for SD1.5/2.x (3 downsamples, latent /8); fix the
  inaccurate "SDXL requires latents divisible by 8" comment.
- inpainting_minimal_inference: remove unused denoise() — the actual loop is
  inlined in main().
- BaseDataset.random_mask: drop unused ratio / mask_full_image parameters and
  the dead cloud_mask import.
- Rename per-prompt --img directive to --i for consistency with Musubi Tuner
  (which already uses --i alongside the shared w/h/l/d/s directives).
- When --train_inpainting is set but the prompt has no --i, warn and skip the
  sample instead of falling through to the standard pipeline (which would
  crash the 9-channel UNet on 4-channel input). Matches the existing
  missing-file skip behaviour.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

The lpw inpaint encode path had two related dtype issues that surfaced
when training-time sample generation hit the VAE inside the
accelerator.autocast() block in train_util.sample_image_inference:

1. Inputs were cast to the UNet dtype (`dtype = unet.dtype`) rather than
   `self.vae.dtype`. With `--no_half_vae` or fp32 VAE this fed the
   wrong-precision tensor into the encode.
2. More importantly, even after fixing (1), fp16 autocast forces conv
   kernels to fp16 regardless of input/weight dtype, and SDXL VAE
   produces NaN in fp16 — so `vae.encode` inside autocast NaN'd out
   even with fp32 weights and fp32 inputs.

Cast inpaint inputs to `self.vae.dtype` and wrap the encode in
`torch.autocast(..., enabled=False)`. Mirrors how the existing
`decode_latents()` runs outside the caller's autocast block.

Verified on SDXL with `--mixed_precision fp16 --no_half_vae` (was NaN,
now produces valid samples) and `--mixed_precision bf16` (no
regression). Same change applied to the SD1.5 lpw pipeline; SD1.5 VAE
is more fp16-tolerant so the bug was latent there.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Inpainting cleanup: misc fixes following PR #2309 review

Fix NaN in inpainting sample generation under fp16 autocast

…distinction

- Sampling/inference uses wobbly_ellipse_mask, not the training-time
  random_mask mixture. Note this in the "Sample images" section to
  prevent confusion.
- Add a Notes bullet distinguishing --train_inpainting from the
  unrelated --alpha_mask loss-mask feature.

The dataset-config note only mentioned cache_latents. cache_latents_to_disk
auto-enables cache_latents (train_util.py:4532), so it is equally
incompatible with --train_inpainting; spell that out for readers.

Remove unused import.

Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>

Remove unused import.

Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>

…l training support