Scenario Thumbnails

Worked example: a boss-fight devlog cover

The user supplies the approved moss-covered stone boss reference, requests one illustrated 1280x720 YouTube thumbnail and the exact headline "ONE HIT LEFT". The boss has two arms, one amber eye and no mouth. The chosen composition is the boss on the right, a small shield-bearing player below, and copy on the left. The authorized ceiling covers one generation and a bounded repair.

1. Inspect the reference and record the identity invariants. Upload it via scenario if it is not already an asset. Preserve its approved illustration style and do not substitute a photographed person.
2. recommend for txt2img with that brief: this is a new composition using a character reference, not a restyle of the reference's layout. Read model_schema_get on a suitable reference-taking member. Check supported ratio, reference count/cardinality, prompt limit and size before selecting the payload. Pass the reference as the schema's actual field, scalar or array as required; do not assume a source-image/strength editing contract. Use img2img when the task is to edit or restyle an existing composition.
3. Prompt the approved boss and action with a clear left-side copy region. Keep the plate free of text and duplicate bosses. Price the payload with model_run and dry_run: true, then submit within the ceiling and wait on its existing job id using jobs_wait.
4. Inspect the result against the boss reference before adding type. A changed eye count or silhouette needs a focused correction; adding text does not repair identity. Use the accepted plate for an edit when appropriate, preserving approved features rather than restarting the concept.
5. Render the supplied headline as a transparent PNG card through scenario-text-overlay, then composite it over the plate with model_scenario-compose-image. That id is fixed rather than discovered because it is Scenario's single deterministic image compositor; it takes image layers only, so the words never pass through a model. Inspect the finished image, not just a browser preview. Land the exact output size through scenario-image/scenario-formats and read the returned dimensions.
6. Review at 1280x720 and as a roughly 120-pixel-wide preview. The small preview is a stress test, not a universal platform standard. Check that the boss and shield still read, the headline is legible, and neither copy nor cropping covers the defining eye. Inspect the actual reduced image; requesting high resolution alone is not this test.
7. Deliver the verified final image through asset_display and asset_download. Keep the clean plate, exact text, font/layout settings and source/output identifiers so a copy change can reuse the art. State any unresolved defect rather than silently marking a failed variant complete.

Review, revisions and common mistakes

For multiple outputs, give each concept or variant a stable label and retain its source roles, payload, job id and QA result. Pending outputs are waited on, not resubmitted. Retry only a failing output within the remaining ceiling. On revision, map the user's note to composition, identity, copy or export; reuse accepted layers and change only the requested part. A new headline normally needs a new overlay, not another image generation.

• Treating a style example as the character source swaps identity without permission.
• A sharp large image can still have an unreadable hierarchy at feed size.
• Longer copy needs a layout decision, not automatic truncation or altered wording.
• A logo is an asset to preserve, not a word to ask a model to spell.
• Decorative UI, rewards and mechanics that were never supplied misrepresent the game.
• Readability and reference fidelity are observable; click-through improvement is not established by a thumbnail review.