Scenario Image Editing

Upscaling is a model family, not a knob

Upscalers are img2img models, many taking no prompt at all: discover with filters={"tags": ["image-upscale"]} (13 hits at authoring time). Fidelity upscalers (Topaz, Recraft Crisp) sharpen and enlarge what exists, the pick when output must stay on-model; creative ones (Magnific Creative, the Clarity pair) carry a creativity dial that invents detail and can redraw fine features, so compare against the source. Sizing comes only from model_schema_get and varies per model: a factor, a target resolution or megapixels, or just image with no dial (2x to 16x and 4K to 8K ceilings were typical, not bounds). Cost follows output pixels: dry_run=true, top-level on model_run and never inside parameters, prices the exact size before a batch. An upscale can also outrun model_run's wait budget where the effects above never do: a modest one still returns status: "success" inline, a larger one returns status: "in_progress" and a job id for jobs_wait, re-called with any returned pending_job_ids as job_ids. Purpose-built variants protect seams on tileable textures and continuity on 360 panoramas: see scenario-textures and scenario-skyboxes.

Expanding a canvas is four different tools

• Reframe (model_scenario-gemini-reframe): an aspectRatio enum plus a resolution tier (1K, 2K, 4K), so exact pixels are out of reach; optional prompt. The enum is approximate: 4:5 came back 1856x2304 (29:36), so a true ratio needs a Resize Image pass after it (fit: "cover", so it crops the sliver instead of squashing).
• Smart Reframe (model_scenario-smart-reframe): width and height are required and exact, and it protects on-image text, brand marks and palette. textDensity: "DENSE" costs substantially more.
• Photoroom Expand: exact outputWidth and outputHeight up to 4096, plus a seed.
• Photoroom Uncrop: rebuilds a subject the frame edge cut off.

Resize Image (model_scenario-resize-image, a fixed first-party id: Scenario's single exact-dimension resize tool, so discovery would only re-derive it) is none of these: it scales pixels to width and height (one alone keeps the ratio) and never invents canvas; fit decides what an off-ratio box does (contain, the default, may land smaller than the box, stretch distorts, cover center-crops to exact dimensions).

Both reframes recompose generatively rather than filling canvas, and at tens of times an effect's price they are the chain's expensive step: dry_run them. They re-render, so reframe first and grade after: a grade or grain pass beforehand comes back partly reinterpreted.

Cardinality

Effects take a scalar image. Resize Image (images, max 10) and Grid Maker (images, max 100) are array: true, where a bare id is silently dropped and the run succeeds having ignored it.

Worked example: grade a key art, then export at size

1. upload_asset the file, then upload_asset_complete unless it went inline under ~100KB.
2. Reshape first: Resize Image to the delivery size, with images as an array even for the one file.
3. model_schema_get on model_scenario-postprocessing-lut (a fixed first-party id: Scenario's single deterministic LUT tool, so discovery would only re-derive it), pick a lutStyle from its enum, price it with model_run and dry_run=true, then run it with lutIntensity near 0.6 for a restrained grade.
4. Grain last, on that output, so its texture is sized for the shipping frame.
5. asset_display to review, asset_download to save.

Common mistakes

• Reaching for an image_edit MCP tool, or for local ImageMagick or Pillow: the surface is model_run on tool models.
• Prompting an effect ("more posterized"): they read numbers only. Reframe and the layer extractors take text.
• Slicing or extracting layers before grading: both return one asset per piece, so every piece then needs its own run.
• Saving a traced SVG with asset_download: it comes back rasterized, so take the stored file from asset_get's url with curl -L.