Scenario texture and material workflows

Overview

Scenario generates seamless, tileable textures from text or reference images, upscales them without breaking the tile, and applies PBR materials when texturing 3D meshes. Model availability differs per team and evolves, so always discover models at run time instead of hardcoding IDs: recommend with the need in the user's own words for a capability, search for a member known by name.

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

Step	Tool	Notes
Find texture models	recommend with the need in the user's own words	capability is txt2img or img2img with the texture in prompt (the catalog tags texture models txt2img_texture, img2img_texture); search for a name
Inspect inputs	model_schema_get	Always call before model_run
Generate	model_run	dry_run=true prices a batch first
Wait	jobs_wait	Re-call with pending_job_ids
View and save	asset_display, then asset_download	Download for engine import
Upscale	model_run on a texture upscaler	2x to 8x, tiling preserved

What live search confirmed at authoring time (examples to re-discover, not constants)

• Seamless generation: model_scenario-texture (Scenario Texture) takes a prompt (a tileable hint is appended automatically), width/height from 16 to 3840 in multiples of 16, quality, seed, up to 10 referenceImages for style, and eraseSeam (off by default) with overlap/featherRadius to inpaint away both seam axes.
• Themed texture LoRAs (Flux.1 LoRA, tag sc:texture): floors, marble, concrete, stone walls, wood boards, brick, terracotta, hand-painted, cybernetic, realistic textures. They expose dedicated texture capabilities (txt2img_texture, img2img_texture, controlnet_texture).
• Tiling-safe upscaling: model_sc-upscale-flux-texture (Scenario Texture Upscale), upscaleFactor 2 (the minimum) to 8, presets precise/balanced/creative riding the same strength and controlnet sliders the schema exposes. It re-renders, so expect small tonal drift: precise with low strength minimizes it, and the result is compared against the source before shipping.
• Material-look conversion: model_sc-texture-converter (Texture Converter) turns a flat image into a surface material using raised, shiny, polished, angular sliders and an invert relief toggle.
• PBR maps come from two different routes. For a flat texture, 2D map extractors (recommend, capability="img2img", with the user's words about extracting maps from a flat texture; search only for one already known by name) return a full set in one img2img call: model_patina (PATINA Image to Maps) outputs base color, normal, roughness, metalness, and height, each as its own image asset whose metadata.type names it (texture-albedo, texture-normal, texture-smoothness, texture-metallic, texture-height) in the order requested; the roughness request comes back labeled smoothness, so confirm the convention before wiring it. model_patina-material tiles a PBR material straight from a prompt. For a mesh, 3D texturing and image-to-3D models emit the maps instead (Tripo 3.0 Texturing, Tencent Texture Edit, Meshy 7 Retexture); enable the PBR toggle found via model_schema_get. Retexturing a full mesh is a 3D-to-3D pipeline: see scenario-3d, and scenario-meshy for the Meshy retexture contract.

Worked example: seamless brick, iterated then upscaled

1. recommend with the user's own words as prompt ("seamless tileable weathered brick"), pick the seamless generator (e.g. model_scenario-texture).
2. model_schema_get model_id="model_scenario-texture".
3. model_run with parameters={"prompt": "weathered red brick wall, moss in the mortar joints", "width": 1024, "height": 1024, "eraseSeam": true, "seed": 42}. For a themed pack, price the batch with dry_run=true first, then launch the runs with wait=false.
4. jobs_wait with job_ids=["job_..."] (the ids returned by model_run). A ~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 the output asset.
5. Iterate: rerun with the same seed and an edited prompt, or add referenceImages=["asset_..."] to lock a style.
6. Upscale: recommend with capability="img2img" and the user's own words as prompt ("upscale this texture without breaking the tile"), hold the pick to tiling preservation (a generic upscaler breaks the repeat), model_schema_get it, then model_run with the image and the factor and preset fields the schema names (at authoring time the texture upscaler took upscaleFactor: 2 and preset: "precise"), raising the factor only deliberately: cost follows output pixels, and this leg can only be dry_run-priced once its input asset exists, so a two-step chain is never priced up front.
7. Verify tiling at no cost: the saved PNG's left column against its right and top row against bottom should differ no more than neighboring interior columns do, and a 2x2 self-tile proof sheet shows any seam instantly.
8. asset_download the final asset for engine import.

Common mistakes

• Skipping model_schema_get: parameter names differ per model and most models reject an empty payload.
• Using a generic upscaler on a tileable texture: it breaks the repeat at the seams. Use the texture-specific upscaler, which preserves tiling.
• Generating huge sizes directly: generate near 1024, then upscale 2x to 8x. Generation dimensions cap at 3840.
• Routing a flat texture through a 3D texturing model just to get PBR maps: the 2D map extractors do that in one img2img call, and the 3D models are for meshes.
• Ignoring engine sizing: engines expect square power-of-two textures (the seamless generator defaults to 1024x1024, 1:1). The generator accepts any multiple of 16, so choose 1024 or 2048 deliberately.
• Hardcoding model IDs: availability differs per team. Re-discover each session.