Inputs
Typed, not implied
image, mesh, cad, video, audio, text, metadata. A job is only offered to an engine whose input types match what it holds.
Registry
The registry is the contract between the queue and the machines. Every adapter declares what it consumes, what it produces and which runtime it needs; the studio probes each configuration before it will route a job. Where two backends exist for the same weights, both are listed.
| Engine | Flow | Runtime | Configs | State |
|---|---|---|---|---|
| env_probe | metadata → metadata | linux-process | 1 · 1 bound | bound |
| glbcoder_text_to_glb | text → mesh | llama.cpp build-vk (Vulkan), llama-server --jinja, one stream, plus headless Chromium (ANGLE/SwiftShader) to press the page's Export GLB button; loopback only | 1 · 1 bound | bound |
| skintokens_tokenrig_autoregressive | mesh → mesh | python 3.11 + torch 2.14.1+rocm7.2 (gfx1151) + transformers 4.57.1 + bpy 4.2, SDPA instead of flash-attn | 1 · 1 bound | bound |
| trellis_cpp_image_to_3d | image → mesh | trellis.cpp 0.8.1 prebuilt Vulkan (ggml 0.24.0), 10 GGUF weights 16 GB | 2 · 1 bound | ready |
| trellis_cpp_image_to_3d_rocm | image → mesh | trellis.cpp 0.8.1 prebuilt ROCm (ggml 0.24.0 hip), 10 GGUF weights 16 GB | 2 · 2 bound | ready |
| zero_to_cad_qwen3vl_2b | image → cad | python3.13 + torch 2.15 cpu + transformers 5.22 | 2 · 1 bound | bound |
| frames_to_video | image → video | ffmpeg 9.0.2 + python3 — no model, no GPU, no checkpoint | 1 · 1 bound | ready |
| text_to_music | text → audio | sa3.cpp sa3-generate (Vulkan/ggml), Stable Audio 3 medium GGUF stack | 1 · 1 bound | blocked |
Snapshot: — · — configurations
registered, — bound · engines present:
—/—.
Rows are rendered from the live probe, not typed by hand: unbound configurations stay listed
and refuse work rather than failing mid-job.
Text → image is served outside the adapter registry by ComfyUI (Qwen-Image 2.1) and is not
counted above.
Contract
A lane is only routable once every field below is known. Anything missing keeps it out of the queue.
Inputs
image, mesh, cad, video, audio, text, metadata. A job is only offered to an engine whose input types match what it holds.
Runtime
Binary, version, backend and weight set are recorded per configuration, so a result can be reproduced on the same stack months later.
State
A configuration that cannot resolve its weights or device reports unbound at probe time. The console shows it; the queue skips it.
If you already have a step that eats hours — retopo, rigging, plates, stems — that is the shape of thing this studio is for.