scenario-skyboxes

Scenario Skyboxes and 360 Panoramas

Overview

Scenario hosts dedicated skybox models that produce seamless equirectangular 360 panoramas, plus a seam-preserving skybox upscaler. Always generate with a skybox-specific model rather than a generic image model: these enforce the seam continuity and pole geometry that ordinary text-to-image output lacks.

Connection and the core generation loop: see the scenario skill in this repo. 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

Model IDs and the parameter facts below were live search hits at authoring time. Re-discover them each time: availability differs per team and evolves.

• model_scenario-skybox-flux: text to 360 panorama, a style preset enum (count drifts; read it from the schema), automatic seam and pole correction, optional reference image with a strength slider.
• model_scenario-skybox-gpt: text to 360 panorama guided by up to 10 reference images, quality presets, width and height up to 3840 px.
• model_hunyuan-world-image-to-skybox: one photo of a place to a seamless 360 skybox.
• model_sc-upscale-flux-skybox: 2x to 8x skybox upscale that preserves the seamless wrap.

Worked example: generate, iterate, export

Example: a stylized forest skybox for a game scene.

1. search with target="models", query="skybox", public=true. Pick a text-to-skybox model, for example model_scenario-skybox-flux.
2. model_schema_get for that model. Expect fields like prompt, style, negativePrompt, image, strength, numOutputs, geometryEnforcement, seed. negativePrompt is inert unless negativePromptStrength is above 0.
3. model_run with parameters={"prompt": "ancient pine forest at dawn, mist between trunks, god rays", "style": "cinematic", "numOutputs": 2}. Take style preset names from the schema response. dry_run=true first prices the run; launch with wait=false.
4. jobs_wait with job_ids=[the returned job_id]. Its ~180s timeout is not an error: re-call with the returned pending_job_ids as job_ids, never a second model_run. Then asset_display each output.
5. Iterate on mood: copy the seed from the best result and change only style (cinematic, oil-painting, cyberpunk, and more). To keep composition while shifting look, pass the favorite as image with low strength (0.2 to 0.4). To steer mood from concept art instead, switch to model_scenario-skybox-gpt and pass referenceImages.
6. Export: run model_sc-upscale-flux-skybox with image=asset_id and the smallest upscaleFactor that reaches target (the upscale can cost several times the generation, and its dry_run can only run once the input asset exists, so the chain cannot be priced up front). baseModel defaults to FLUX.1-dev (stylized); a Krea-based realism option exists, and strength defaults to 0.6, which invents detail: lower it when the goal is the same panorama at higher resolution. Read the exact allowed values from model_schema_get before switching. Then asset_download the final asset.
7. Verify the seam at no cost: compare the saved PNG's leftmost and rightmost pixel columns; on a seamless panorama they differ by near zero while columns a quarter-turn apart differ by an order of magnitude more.

Engine format notes: Skybox Flux outputs equirectangular panoramas; keep the default sizing (1536x768 at authoring time) and read the real dimensions off the returned asset rather than assuming them. Skybox GPT's catalog lists equirectangular 2:1 plus cubemap strip 6:1 and cubemap cross 4:3 layouts, but its schema exposes only width and height, so confirm the layout contract with model_schema_get before relying on a cubemap layout. Beyond flat backdrops, the same search surfaces model_hunyuan-world-skybox-to-splat, and a separate search (query="world") finds the Marble world models; both turn a finished panorama into a navigable 3D Gaussian splat scene, the pipeline the scenario-3d-worlds skill teaches end to end.

Common mistakes

• Prompting a generic image model for a "360 panorama": edges will not wrap and poles smear. Use a dedicated skybox model.
• Upscaling with a generic upscaler: it breaks continuity at the wrap seam. Use the skybox upscaler.
• Hardcoding model IDs in scripts or docs: re-discover with search; the catalog changes.
• Fighting seam or pole distortion through prompt wording on Skybox Flux: raise geometryEnforcement above 0 instead, and only when distortion is actually visible.
• Requesting a non 2:1 width to height ratio on Skybox GPT while expecting equirectangular output: keep 2:1 (for example 2048x1024) for correct 360 viewing.
• Skipping model_schema_get: skybox models carry model-specific fields (style, geometryEnforcement, quality) that generic assumptions miss.