feat(generate): restore model-family tuning, Krea-2 style reference and output filters

[joshistoast]

Summary

Model-family tuning controls and two output options from the legacy frontend were missing from webv2 Generate and Canvas.

• FLUX.1 DyPE: preset (off, auto, area, 4K optimized, manual) with scale and exponent under Manual, sent to the FLUX denoise node only when set.
• Z-Image: an optional timestep shift override and the seed-variance enhancer (strength and randomized channels), inserted between the positive conditioning and its collector exactly as legacy wired it.
• Qwen-Image: timestep shift override and encoder quantization (none, int8, nf4) applied to both text encoders.
• Krea-2 style reference: a new reference-image kind with a 0–2 style strength, compiled to krea2_style_reference fed by the (cropped) image and the model VAE and connected to denoising; strength 0 bypasses it. Krea-2 now advertises one reference image on the backend.
• NSFW checker and invisible watermark: two project settings (Settings → Project, off by default) that chain img_nsfw / img_watermark after the final image of every built-in Generate and Canvas family, keeping the output flags and metadata on the last node.

Every control appears only for its model family; graphs at defaults are structurally unchanged. New settings are emitted in metadata and restored by Recall All and Remix.

Related Issues / Discussions

Follows up the legacy parity audit behind #9803.

QA Instructions

From invokeai/frontend/webv2 (Node 22.14, pnpm 10.12): pnpm check:release passes. Python: pytest tests/backend/architectures tests/app/invocations/test_frontend_graph_node_types.py passes.

The mock backend's representative profile has only SD 1.x main models, so the per-family controls are covered by browser tests (GenerateRenderSection, GenerateAdvancedFields) and compile tests. On the mock: open a project → Settings → Project: toggle NSFW checker and Invisible watermark; the Generate graph preview then ends … → img_nsfw → img_watermark. With a real install: select a FLUX.1 model → Render → DyPE = Manual shows λs/λt; Qwen-Image → Advanced → Encoder quantization; Z-Image → Shift override and Seed variance; Krea-2 → add a reference image → Style strength. Generate, then Recall All restores each.

Review

No material findings. Reviewers confirmed every field against the backend invocations and the legacy builders, including that DyPE scale/exponent are read only under the manual preset. Follow-ups applied from review: tests for the shift override being turned off, watermark-only Canvas outpaint to a staging destination, the project setting flowing through the normal invoke path, and recall of a style reference whose image is missing; a single "Shift" label; docs naming.

Limitation: core_metadata at defaults now records z_image_shift: 'auto', qwen_image_shift: 'auto', qwen_image_quantization: 'none' and the seed-variance keys, matching legacy; execution nodes are unchanged.

Compatibility / Rollout

New generate settings and the two project settings are optional with defaults and normalize on load; no project-document migration. The Krea-2 architecture definition now reports max_reference_images=1 (capabilities fixture regenerated). No API changes.

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)

🤖 Generated with Claude Code

---

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