Summary

Refs #10695. This adds four Images-page items from the settings gap list.

1. Paste generation parameters

• Paste entry points:
  • Preset menu: a new "Paste generation parameters" entry opens a text box.
  • Prompt box: pasting a whole AUTOMATIC1111 / Forge / Civitai block (prompt, Negative prompt:, and a Steps: line with a seed, sampler or CFG) into the prompt box fills the form instead of adding the text to the prompt.
  • Unsloth recipes: pasting an Unsloth JSON recipe also works.
• One parser: paste and image drop share one parser (parseA1111Parameters in lib/image-metadata.ts) and one restore path (applyImportedRecipe). That path also produces the "not applied" toast.
• Format: the parser follows modules/infotext_utils.py parse_generation_parameters:
  • The last line is the settings line when it has Steps:, or 3+ fields including a known key (the re_param rule, :254-257).
  • Quoted values are unquoted.
  • JSON values (Civitai's Civitai resources: [...], Hashes: {...}) stay whole instead of being split at their commas.
• New keys:
  • Clip skip (A1111 numbering)
  • CFG Rescale / RescaleCFG / Forge's latent_modifier_rescale_cfg_phi
  • Variation seed / Variation seed strength
  • Hires *
  • Mask blur, Inpaint area, Masked area padding
• LoRA tags: <lora:name:weight> tags (also te= / unet= and <lyco:...>) are removed from the prompt. Each one selects the installed local LoRA with the same file name, at that weight clamped to 0-2. Tags with no match are reported as "LoRA x (not installed)". Other tags such as <hypernet:...> are removed from the prompt and reported too.
• Not applied: settings this build cannot apply are listed in the toast instead of being dropped without a word. Examples: Clip skip above 2, CFG rescale, variation seed, hires fix, VAE, ADetailer keys, and samplers the loaded model does not offer.
• Clip skip 1 and 2 are not reported. On SDXL, A1111 and Forge render both from the penultimate layer, and that is already what Studio renders.
• Ignored on purpose: file-identity keys such as Model hash, Lora hashes and Version are read and ignored.

2. Inpaint mask blur and "Only masked"

• Backend: three optional request fields.
  • mask_blur (0-64 px): a Gaussian blur of the mask. This is what diffusers VaeImageProcessor.blur does.
  • inpaint_only_masked
  • inpaint_padding (0-256, 32 when left out, the A1111 default). It maps to diffusers padding_mask_crop.
• Gating:
  • The fields apply only in the inpaint workflow.
  • "Only masked" is accepted only when the inpaint pipeline's __call__ takes padding_mask_crop. Otherwise the request gets a 400.
  • A new status flag supports_inpaint_only_masked reports this. It is False on the native and MLX engines.
  • Applied values are saved in the gallery record and written to the image's parameters text, so they restore.
• Frontend: a "Mask blur" slider in Inpaint. The "Only masked" switch and its Padding slider appear only when the loaded model supports them. At their defaults the controls send nothing.

3. Seed buttons

"Random" (clears the Seed field: a new random seed each run) and "Reuse last seed" sit next to Seed. Reuse fills in the base seed of this session's last run. Before any run this session, it uses the newest gallery record's seed.

4. Steps slider max

• Problem: the slider stopped at 50, but Krea-2-raw's default is 52 steps.
• Fix: the max is now min(100, max(50, 2 x model default, current value)). 100 is the request's limit.
• Effect: models with a default of 25 or less keep the 50 max. Imported or restored values above 50 can now be shown.

Testing

• Frontend unit tests: new tests/image-paste-parameters.test.ts and tests/image-seed-inpaint-steps.test.ts.
  • Before (origin/main): both fail at import, because the parser, seed, inpaint and steps helpers do not exist there.
  • After: all pass.
  • Base parser on a Civitai block: it leaves <lora:...> in the prompt and drops Clip skip, CFG Rescale and Variation seed with no message.
• Changed tests in tests/image-metadata.test.ts:
  • It used cfg_rescale as its example of an unknown key. That key is now listed by label, so the example moved to tiling.
  • One page-source anchor was updated for the new not-applied list.
• Full frontend suite: node --experimental-strip-types --test "tests/**/*.test.ts": 13244 tests, 0 failures. node scripts/typecheck.mjs --tests is clean.
• ESLint: images-page.tsx has 37 errors on both base and head, so this change adds none.
• New backend tests:
  • tests/test_diffusion_inpaint_mask.py (CPU, 14 tests): request bounds, blur, padding, the status flag, and the route passing fields only when set.
  • tests/test_diffusion_inpaint_mask_sdxl.py (5 tests, through Studio's real DiffusionBackend with a tiny SDXL inpaint pipeline on GPU), which checks that:
    • Leaving the fields out gives the same pixels as setting them to None, 0 or False, and no padding_mask_crop is passed.
    • mask_blur=8 changes the mask the pipeline receives, and the output.
    • "Only masked" passes padding_mask_crop, keeps the source size, and leaves pixels away from the mask unchanged.
    • A pipeline without padding_mask_crop is refused.
• Bit-identity against origin/main: a default inpaint render with seeds 7 and 8 gives the same pixel hashes on base and head (maximum pixel difference 0).
• Existing related backend tests: diffusion backend, routes, sampler, SDXL, ControlNet, image gallery, request validation and the generate-image tool. Head: 971 passed, 0 failed. In a full origin/main checkout the same files pass too.

Not covered

• The only inpaint pipeline actually rendered was the tiny SDXL one. FLUX / Qwen-Image / Z-Image / FLUX.2-klein inpaint also report padding_mask_crop support but were not rendered.
• No HTTP round trip through the generate route. The route helper is covered by unit tests only.
• There is no browser screenshot of the new controls.
• Clip skip, CFG rescale and variation seed are parsed and reported as not applied. They become applied once the request accepts those fields.

Review follow-ups

• An A1111 block with Mask blur, Inpaint area or masked padding next to Denoising strength now imports as inpaint, not img2img.
• Restoring an inpaint record resets mask blur, Only masked and padding to their defaults when the record omits them.
• A run whose response was lost still records its base seed for Reuse last seed.
• Reuse last seed is disabled, rather than offering an older pinned image's seed, when the loaded page is all pinned and more images exist.
• An imported Only masked area is listed under "Not applied" when the loaded pipeline cannot crop.
• The external-connection route refuses mask_blur, inpaint_only_masked and inpaint_padding with a 400, like the other conditioning fields.
• Mask blur now feathers the result: the diffusers inpaint mask processors binarize mask_image, so with Only masked off the render is pasted back over the source through the blurred mask (A1111's apply_overlay). With Only masked on, padding_mask_crop already overlays with the blurred mask, so it is not applied twice.
• Pasting parameters with <lora:...> tags before a LoRA-capable model is loaded matches the tags against the full local LoRA list instead of reporting installed adapters as missing.
• The Reuse last seed button uses lucide Recycle instead of a HugeIcons barrel import, which kept the eager startup bundle under its budget (main + PR: 2023.3 KB transfer, over the 2023.0 KB budget; with this change 2022.9 KB). The screenshots below were taken before this swap, so they show the old icon.

Screenshots: before (merge base dff0fb917) vs after (44cf94475)

Two separate install.sh --local builds of Studio, one from each commit. Both load the same model: a tiny SDXL pipeline with random weights (hf-internal-testing/tiny-stable-diffusion-xl-pipe), run as family sdxl from a local folder whose name contains krea-2-raw. Studio looks up a model's default steps by its name, so that folder name gives a resident default of 52 steps on both sides (the backend also reports generation_defaults.steps = 52). That shows the Krea-2-raw case without downloading Krea-2-raw itself. The noise image in the gallery is what random weights produce.

Steps slider at a 52-step model default. Before, the slider stops at 50 and the thumb sits pinned at the end while the field reads 52. After, the slider goes up to 100 (aria-valuemax 50 -> 100) and the thumb sits at 52.

Seed buttons. Before, there is only the Seed field. After, Random and Reuse last seed sit next to it. This shot was taken after one generation with seed 424242. Clicking Reuse put 424242 in the field, and the tooltip reads "Reuse last seed (424242)".

Preset menu. After, the menu has a new "Paste generation parameters" entry under "Load settings from an image".

Pasting a Civitai-style A1111 block. The block includes <lora:sloth_style_v2:0.8>, Clip skip 2, CFG Rescale 0.7 and Variation seed 42. Before, pasting it into the prompt box drops the whole block into the prompt as plain text. After, the block goes into the Paste generation parameters dialog.

The form after the paste. Before, the prompt holds the raw block, including the LoRA tag. Steps stays at 52, the size stays 1024x1024 and the seed stays empty. After:

• The prompt has the LoRA tag removed.
• The negative prompt is filled in.
• Steps 30, CFG 6.5, sampler DPM++ 2M / Karras, seed 1234567890 and 832x1216 are all applied.
• The toast lists what was not applied: "CFG rescale 0.7, variation seed 42 (strength 0.3), LoRA sloth_style_v2 (not installed)". Clip skip 2 is not listed, which is intended: SDXL already renders from the penultimate layer.

Inpaint with a source image. Before, only Strength is shown. After, there is a Mask blur slider (0-64, default 0) and an Only masked switch, which appears because supports_inpaint_only_masked is true for SDXL. Turning the switch on shows a Padding slider (0-256, default 32).