feat(controlnet): add Qwen Image Fun-ControlNet-Union with canvas control layers

[Pfannkuchensack]

Summary

Qwen-Image had no ControlNet. This adds alibaba-pai's Fun-ControlNet-Union (Canny, HED, depth, pose, MLSD, scribble, gray, plus inpainting) as a Qwen Image ControlNet: a model config/loader, the qwen_image_control node, control and inpaint conditioning in qwen_image_denoise (hints injected through forward hooks on the diffusers transformer), a canvas control-layer kind with inpaint/outpaint feeding, a grayscale filter, a starter model and docs. The denoise now reserves working memory for its activations and materialized attention scores.

Related Issues / Discussions

Part of the Fun-ControlNet series (Qwen-Image 2.1, FLUX.2 dev, InstantX and the Z-Image migration follow).

QA Instructions

• Parity on the real model (RTX 4090, Qwen-Image 2512 Q4_K_M + Union 2602, 1024²): the hooked forward is torch.equal to an independent reference through diffusers' controlnet_block_samples path; scale 0 and inactive hooks equal the plain forward.
• API runs: control at 0.85 and a 0–50 % window steer the composition (control-edge coverage 0.66–0.93 vs 0.23–0.55 without); pure inpaint keeps everything outside the mask (mean |Δ| ≈ 3) and regenerates inside; combined mode works; a missing VAE is refused with a clear message.
• Canvas: a control layer defaults to the kind and the union model, Canny filter applied, generation follows the edges; a second Qwen Image ControlNet layer blocks Invoke with the limit reason; Edit models are refused.
• uv run --no-sync pytest -n 4 tests/backend/fun_controlnet tests/backend/model_manager tests/backend/architectures tests/model_identification tests/backend/z_image tests/app/invocations: 5316 passed. Ruff clean.
• pnpm lint clean; pnpm test: 10486 passed, 3 known German-locale failures; browser tests for the layer settings pass.

Review

Resolved: a materialized-score term was added to the working-memory estimate for builds without fused masked attention; the control chain runs in the branch dtype and hands hints back in the base dtype; compiled canvas graphs (txt2img, inpaint, outpaint) are validated against the backend invocations; docs cover the grayscale filter and tiled control encodes. Remaining: the working-memory constants are estimates pending calibration; on 16 GB cards the 3.5 GB branch streams beside the transformer; control is refused for Edit models and with reference latents.

Compatibility / Rollout

New node and config class; qwen_image_denoise 1.1.0 adds optional control and vae inputs. openapi.json/schema.ts and the capabilities fixture are regenerated. Canvas documents gain an additive adapter kind (no schema bump). No database migration.

Checklist

• The PR has a short but descriptive title, suitable for a changelog
• Meaningful regression coverage added / updated where needed; obsolete tests/code removed
• Persisted-state and API changes include required migrations / compatibility validation
• Relevant performance/efficiency opportunities considered; material claims have evidence
• Material review findings resolved and relevant checks rerun
• Documentation added / updated (if applicable)
• Updated What's New copy (if doing a release after this PR)

---

^{Need help on this PR? Tag @codesmith-bot with what you need.}