feat(anima): Anima Controlnets + Inpainting Adapter

[kappacommit]

Summary

Adds support for kohya-ss's ControlNet-LLLite adapters for Anima — small add-on models (8–66 MB) that bring two user-facing improvements:

• Better inpainting and outpainting. With the new optional Inpaint Adapter (Advanced settings), the model is conditioned on the surrounding image and the results blend in cleanly. In fixed-seed side-by-side tests, seam artifacts dropped by ~60% and the composition breaks disappeared.
• Control layers for Anima. Sketch/scribble, lineart, depth, and pose adapters can now guide Anima generations from a drawing or reference image — the same Control Layer workflow as other model families.

All variants are available as one-click starter model installs, work with existing Anima setups, and add no measurable generation time. The inpaint adapter applies automatically when inpainting/outpainting on Canvas; control adapters are picked per control layer.

Under the hood: a new backend module (ported with attribution from kohya's Apache-2.0 reference implementation), model-manager registration that routes inpaint vs. control variants to the right pickers, an anima_lllite node, and canvas graph wiring for both surfaces.

Related Issues / Discussions

N/A — integrates https://huggingface.co/kohya-ss/Anima-LLLite (reference implementations: kohya-ss/ComfyUI-Anima-LLLite, kohya-ss/sd-scripts).

QA Instructions

Requires the Anima starter models (Base 1.0, Qwen3 0.6B encoder, QwenImage VAE).

1. Inpainting Adapter

• Install it via starter models
• In the advanced section below the VAE/Encoder, select the inpainting adapter
• Do inpainting just as you always would, apply an inpaint mask and go
• Inpainting quality should appear marginally more coherent and logical

2. Controlnets

• Install via starter models
• Use just as you would any other controlnet

Merge Plan

No DB schema changes. Redux changes are additive with defaults (no slice migration needed). The 7 commits are two logical stacks (inpainting adapter, then control layers) and can be reviewed in order.

Checklist

• The PR has a short but descriptive title, suitable for a changelog
• Tests added / updated (if applicable)
• ❗Changes to a redux slice have a corresponding migration (not required — additive fields with defaults, per existing precedent)
• Documentation added / updated (if applicable)
• Updated What's New copy (if doing a release after this PR)

🤖 Generated with Claude Code

[Pfannkuchensack]

Findings

• Medium: invokeai/frontend/web/src/features/nodes/util/graph/generation/buildAnimaGraph.ts:271-289
  The new control_lllite collect-node lifecycle is logic-heavy and entirely untested. The accounting
  addedLLLiteAdapters = hasInpaintAdapter ? 1 : 0 (line 272) is then incremented by
  addAnimaLLLiteControl's return, and the collect node is wired to denoise only if the total is > 0,
  otherwise deleted (lines 284-288). A regression here has two concrete failure modes: (a) an empty
  collect node left wired into denoise.control_lllite -> backend graph/validation error, or (b)
  hasInpaintAdapter counted as 1 while addInpaint/addOutpaint did not actually wire the node ->
  collect kept but under-fed. There is already a graph-builder test file at
  invokeai/frontend/web/src/features/nodes/util/graph/generation/buildAnimaGraph.test.ts, and it
  contains ZERO references to lllite/control_lllite (verified: grep count 0). The inpaint-adapter
  wiring in addInpaint.ts:203-223 / addInpaint.ts:297-317 and addOutpaint.ts:170-190 /
  addOutpaint.ts:319-339 (including the white=inpaint mask inversion via img_lerp min=255/max=0) is
  likewise unexercised.
  To expose this issue, add a test that builds the Anima graph for generationMode 'inpaint' with and
  without params.animaLLLiteModel and asserts: the control_lllite collect node exists and has an edge
  to denoise.control_lllite exactly when an adapter is present, that it is absent/deleted when no
  adapter and no control layers exist, and that the anima_lllite node receives both an image edge and
  an inverted-mask edge.

• Medium: invokeai/app/invocations/anima_denoise.py:304-344 (_build_lllite_cond_image) and
  invokeai/app/invocations/anima_denoise.py:346-359 (_get_lllite_multiplier)
  Both contain branch logic with no automated coverage. _build_lllite_cond_image raises ValueError for
  a 4-channel (inpaint) adapter when mask_name is None (lines 324-328), logs a warning and drops the
  mask for a 3-channel adapter (lines 339-343), and raises for any other channel count (lines 334-338).
  _get_lllite_multiplier gates the weight to 0.0 outside [floor(begintotal), ceil(endtotal)].
  The Anima LLLite backend test module covers from_state_dict, passthrough, bind/restore, multi-adapter
  composition, and _normalize_control_lllite, but neither of these two methods. A regression that, e.g.,
  inverts the channel test or the step-range comparison would silently mis-condition inpainting or
  apply an adapter outside its intended step window with no failing test.
  To expose this issue, add a test that calls _build_lllite_cond_image with a stub model of
  cond_in_channels=4 and mask_name=None and asserts ValueError; with cond_in_channels=3 and a mask
  asserts the returned tensor has 3 channels and a warning was logged; and a test that asserts
  _get_lllite_multiplier returns 0.0 for a step below first_step and above last_step and the field
  weight in between.

The Inpainting is so much better with that. I tested the others Controlnet and had only the problem that the 1.0 Weight is a bit low to show a good effect on the control. I had to use atleast 1.35~1.55.

[kappacommit]

Findings

* Medium: invokeai/frontend/web/src/features/nodes/util/graph/generation/buildAnimaGraph.ts:271-289
  The new control_lllite collect-node lifecycle is logic-heavy and entirely untested. The accounting
  `addedLLLiteAdapters = hasInpaintAdapter ? 1 : 0` (line 272) is then incremented by
  addAnimaLLLiteControl's return, and the collect node is wired to denoise only if the total is > 0,
  otherwise deleted (lines 284-288). A regression here has two concrete failure modes: (a) an empty
  collect node left wired into denoise.control_lllite -> backend graph/validation error, or (b)
  hasInpaintAdapter counted as 1 while addInpaint/addOutpaint did not actually wire the node ->
  collect kept but under-fed. There is already a graph-builder test file at
  invokeai/frontend/web/src/features/nodes/util/graph/generation/buildAnimaGraph.test.ts, and it
  contains ZERO references to lllite/control_lllite (verified: grep count 0). The inpaint-adapter
  wiring in addInpaint.ts:203-223 / addInpaint.ts:297-317 and addOutpaint.ts:170-190 /
  addOutpaint.ts:319-339 (including the white=inpaint mask inversion via img_lerp min=255/max=0) is
  likewise unexercised.
  To expose this issue, add a test that builds the Anima graph for generationMode 'inpaint' with and
  without params.animaLLLiteModel and asserts: the control_lllite collect node exists and has an edge
  to denoise.control_lllite exactly when an adapter is present, that it is absent/deleted when no
  adapter and no control layers exist, and that the anima_lllite node receives both an image edge and
  an inverted-mask edge.

* Medium: invokeai/app/invocations/anima_denoise.py:304-344 (_build_lllite_cond_image) and
  invokeai/app/invocations/anima_denoise.py:346-359 (_get_lllite_multiplier)
  Both contain branch logic with no automated coverage. _build_lllite_cond_image raises ValueError for
  a 4-channel (inpaint) adapter when mask_name is None (lines 324-328), logs a warning and drops the
  mask for a 3-channel adapter (lines 339-343), and raises for any other channel count (lines 334-338).
  _get_lllite_multiplier gates the weight to 0.0 outside [floor(begin_total), ceil(end_total)].
  The Anima LLLite backend test module covers from_state_dict, passthrough, bind/restore, multi-adapter
  composition, and _normalize_control_lllite, but neither of these two methods. A regression that, e.g.,
  inverts the channel test or the step-range comparison would silently mis-condition inpainting or
  apply an adapter outside its intended step window with no failing test.
  To expose this issue, add a test that calls _build_lllite_cond_image with a stub model of
  cond_in_channels=4 and mask_name=None and asserts ValueError; with cond_in_channels=3 and a mask
  asserts the returned tensor has 3 channels and a warning was logged; and a test that asserts
  _get_lllite_multiplier returns 0.0 for a step below first_step and above last_step and the field
  weight in between.

The Inpainting is so much better with that. I tested the others Controlnet and had only the problem that the 1.0 Weight is a bit low to show a good effect on the control. I had to use atleast 1.35~1.55.

updated with new tests based on feedback.

regarding the weak controlnets, unfortunately the model creator themselves calls out that theyre kind of weak due to being trained on the previous version of the Anima (Preview 3). But to my knowledge there are no other controlnets out there, these are still the best available.

[Pfannkuchensack]

Good work.