[Pfannkuchensack]

Summary

The Upscale widget can now use Krea-2 Turbo and Krea-2 Raw main models. The pass works like FLUX.1, Z-Image and Anima: Spandrel enlarges the image, then one img2img pass runs over the whole frame. It encodes and decodes with the Qwen-Image VAE (qwen_image_i2l / qwen_image_l2i, both tiled), the same nodes Generate and Canvas use for Krea-2. No new nodes.

• Components: a single-file or GGUF Krea-2 needs a Qwen3-VL Encoder (the 4B) and a Qwen-Image VAE, both pre-filled. Ideogram 4's 8B encoder is the same model type and is never offered, since krea2_model_loader refuses it. A Diffusers install brings its own.
• CFG: picking the model sets its served default: 1 for Turbo, 5.5 for Raw. The negative prompt is wired only above CFG 1.
• Scheduler: krea2_denoise has no scheduler input, so the field is hidden and the metadata records Euler. Steps keep the widget value for both variants.
• Creativity: krea2_denoise clips denoising_start by step index, like Z-Image and Anima, so the widget translates Creativity to the step nearest FLUX's entry sigma. Two things differ from those nodes:
  • Krea-2 rounds start * steps (Python's half-to-even) where the others floor it, so its start is the step itself, not the middle of the step.
  • Its shift depends on the variant: Turbo's mu is fixed at 1.15, and Raw's follows the frame size.
  • To carry this, the table field indexClippedShift(size) becomes indexClipping: { shift(model, size), rounding }.
• Balanced at 30 steps on a 2048² result starts at step 23 for both variants and runs 7 steps.

Depends on #9887, which holds Raw's mu at 1280×1280 of area. Without it, Raw at 4096² enters at sigma 0.98 and replaces the image.

Related Issues / Discussions

Block 5 of the upscale work. Follows #9825 (Z-Image and Anima). Requires #9887.

Until #9887 is merged, this branch also carries its commit (22be8138d9, the Raw shift hold). Only the second commit (6a38e3f957) belongs to this PR; the branch will be rebased onto main once #9887 lands.

QA Instructions

Automated

• webv2, run one by one:
  • lint (oxfmt, oxlint, tsc --noEmit, architecture): pass.
  • test: 10,196 passed. On this German-locale machine, 3 tests that format numbers with the system locale fail the same way without this change.
  • Browser tests rendering the Upscale and Generate panels: 72 passed.
  • test:performance:architecture (build and browser), test:project-files:browser, test:accessibility:browser (19) and test:fixtures (32): pass.
• test_upscale_index_clipped_starts.py and test_frontend_graph_node_types.py: 889 passed, 3 skipped.
  • The fixture now records Krea-2 Turbo and Raw cases with their variant: Raw at 512/1024/2048/4096, so the hold is crossed.
  • Each case drives the real krea2_denoise._run_diffusion with the real diffusers scheduler up to the img2img blend, and checks that the entry is the step nearest FLUX's.
  • Translating Krea-2 with the floor rule fails 33 cases. A backend without fix(krea2): hold Raw's schedule shift at the end of its fit #9887's hold fails 14.
• The upscale graph contract now covers six architectures: 45 node types and 189 literal values.

End to end (RTX 4090, 1024² source, Balanced, through the Invoke button):

Model	Result	Start	Time
Krea-2 Turbo int8, CFG 1	2048²	23/30, no negative branch	75 s
Krea-2 Raw Q4_K_M GGUF, CFG 5.5	2048²	23/30, negative branch	86 s
Krea-2 Raw Q4_K_M GGUF, CFG 5.5	4096² (4×)	23/30	8:46 min

All three keep the composition and lettering:

• Turbo adds paper grain and painterly stone texture.
• Raw stays closer to the source's colors.
• At 4096², a 100% crop is slightly smoother than the frame.

UI, checked in the browser

• Picking Krea-2 Turbo shows CFG 1, no scheduler field, the Qwen3-VL Encoder picker pre-filled with the 4B (not the 8B), and the Qwen-Image VAE. Raw shows CFG 5.5.
• The output size reads 2,048 × 2,048 for the 1024² source at 2×.

Review

Findings resolved:

• The docs' memory section now covers Krea-2.
• The Python scheduler check ties Krea-2's Euler-only list to the node having no scheduler input.

Remaining:

• The frontend copies the stock Krea-2 shift constants. A Diffusers install samples with its own scheduler/scheduler_config.json. Krea's ships the same numbers. A checkpoint whose config differed would move Raw's starts below 1280×1280 without the contract test noticing.

Compatibility / Rollout

• Merge order: merge fix(krea2): hold Raw's schedule shift at the end of its fit #9887 first. Raw's translated starts above 1280×1280 assume its hold.
• Persisted widget values: one new nullable field, qwen3VlEncoderModel. Older values normalize it to null, and the model sync fills it in.

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.}

[lstein]

CI was failing on all three Python test jobs because SUPPORTED_BASE_COUNTS in tests/app/invocations/test_frontend_graph_node_types.py still pinned the upscale panel at 5 architectures, while the regenerated snapshot now lists 6 (Krea-2 added). This was a trivial fix, so I took the liberty of updating the branch directly (6c1a931: "upscale": 5 → 6).