scenario-seedream

Scenario Seedream Images

Overview

Seedream, ByteDance's image family on Scenario, spans generation, reference-driven editing, and one member that only takes images apart: Layerize splits a finished image into editable layers. The members agree on little else, so discover them with search and read model_schema_get before every run.

Connection and the core loop: see the scenario skill; model-agnostic image work: the scenario-image skill. If a sibling skill named here is missing from your available skills, ask the user to install it (npx skills add scenario-labs/skills --skill <name>); unattended, proceed from tool schemas and flag the gap.

Quick reference

At authoring time (fields and caps are per member, read the live schema):

Member	References	Sizing	Notes
5.0 Pro	referenceImages, up to 10	exact width and height, 672 to 3136 px, step 16	exact in-image text; 3500-char prompt
5.0 Lite	referenceImages, up to 14	width and height, up to 4K	fast and cheap; 2048-char prompt
5.0 Flash	referenceImages, up to 10	exact width and height, 672 to 3136 px, step 16	cheapest; background keeps a reference's alpha
4.5	referenceImages	size (2K, 4K) plus aspectRatio enum	subject-preserving edits; "auto" ratio follows the source
5.0 Pro Layerize	one image, required	size tier: auto, 1K, 1.5K, 2K	splits, never generates; prompt optional
5.0 Flash Layerize	one image, required	size tier: auto, 1K, 1.5K, 2K	same contract minus optimizePromptMode

On the generators, mode follows from the inputs: empty referenceImages is text-to-image, one or more is an edit or a multi-reference generation (with several, the prompt gives each reference a role by position), and the array shape holds even for one asset. Sequence mode (sequentialImageGeneration: "auto" plus maxImages) lets Lite and 4.5 return a related set in one run, input plus generated capped at 15 images; Pro has no sequence fields. Price per image differs widely: a 2048 px square quoted 4 CU on Flash, 6 on Lite, and 18 on Pro at authoring time, and Pro took about two minutes against under one, so iterate on the cheap members and spend Pro on finals; dry_run both before a batch (the estimate prices the run exactly as submitted, a whole sequence included). Pro and Pro Layerize carry optimizePromptMode: the default standard reasons about the prompt first and is slower; fast costs the same and is usually enough when a reference already sets the composition (on Layerize it trades some split fidelity for speed). Flash's schema says width and height apply only when Resolution is Custom, but Flash has no Resolution field: the pixels apply directly.

Keeping a transparent asset transparent

Flash's background: "transparent" preserves alpha through an edit; it does not cut a subject out. It keeps the transparent surround, not translucency inside the subject: glass at alpha 170 came back near-opaque in one authoring-time test, so composite a translucent part in post. The field defaults to opaque, so a restyle that omits it flattens a valid cutout: set it on every run that must stay transparent. It applies only with exactly one reference image that already carries an alpha channel, and is ignored otherwise: a text-only run, a flattened JPEG reference, or two references all come back opaque without an error. So restyle or recolor an existing cutout (a sprite, an icon, a prop) on Flash with the PNG as the one reference, and check the result's alpha on the downloaded file before shipping it, since asset_display composites transparency away: with Pillow (uv run --with pillow when it is missing), Image.open(path).getchannel("A").getextrema() should start at 0 (fully transparent pixels exist), and an image with no A channel was flattened. To get a cutout in the first place, generate on a plain field and run background removal (the scenario-image-editing skill), or split the image with Layerize.

Write the exact copy into the prompt

Pro renders legible in-image text, multi-line layouts and non-Latin scripts included. Quote the exact strings in the prompt instead of paraphrasing them, give each a position and a size rank (headline across the top, date line small at the foot), and keep to a few elements at one or two sizes: letters are drawn, not typeset, so paragraphs and many small labels turn to shapes, and copy past that budget (tour dates, credits) is composited in post from the start. Proofread with asset_display; a word that garbles is spelled out letter by letter after the quoted string on the rerun, and one that still garbles goes to post. To swap copy on a reference, quote the new string, pin its position, and require the original typeface, size and color with everything else unchanged.

Layerize: one image in, an editable stack out

Layerize returns a base layer plus up to 16 transparent PNG cutouts, rebuilding the background behind whatever it lifts. The prompt picks the mode: empty runs a full automatic split; an enumerated list of parts ("Separate this poster into transparent layers: headline, product, shadow, background") cuts better than "all layers"; <bbox>x1 y1 x2 y2</bbox> on a 0 to 1000 grid, origin top left, confines the split to one region. Set size explicitly, since auto inherits the source's tier and the tier moves the price; cost is otherwise flat per run, not per layer. Layers come back cropped to their own bounds with bbox and z-index metadata, not aligned to the source canvas, and there is no PSD export.

Worked example: a poster, then its layers

1. search with target="models", query="seedream", public=true. Prefer the newest non-deprecated hit for the job, e.g. model_bytedance-seedream-5-0-pro (a live hit at authoring time: re-discover each session).
2. model_schema_get with that id: sizing fields, caps, defaults.
3. model_run with that id, dry_run=true, and parameters={"prompt": "Concert poster, teal and cream, screen-print grain. Headline \"MIDNIGHT ORBIT\" across the top, date line \"Nov 14, Union Hall\" small at the foot.", "width": 1600, "height": 2368}; both sizing fields move price.
4. Rerun model_run with wait=false, then jobs_wait with the returned job id, re-called with pending_job_ids on timeout, never a second model_run.
5. asset_display and proofread the rendered text.
6. model_schema_get on the Layerize hit, then model_run with that id and parameters={"image": "<poster asset id>", "prompt": "Separate this poster into transparent layers: the headline text, the date line, and the background. Clean edges, complete transparency.", "size": "2K"}.
7. jobs_wait, then asset_display each layer and asset_download the keepers.

Common mistakes

• Reusing one member's parameter block on another: pixels on Pro and Lite, tier enums on 4.5 and Layerize; pixels sent to an enum field are rejected.
• Passing referenceImages to Layerize or image to the generators: the input field's name and shape differ per member.
• Expecting Layerize layers to overlay the source directly: each is cropped to its own bounds, so reposition with the returned metadata.
• Asking Pro for a sequence: the sequence fields existed on Lite and 4.5 only.
• Moving a 3000-character prompt from Pro to Lite: prompt caps are per member.
• Setting Flash's background to transparent on a text-only run or a flattened reference: it is silently ignored, and the image comes back opaque.