BRP AI Hub ONNX Models
Deployment-oriented ONNX conversions used by BRP Canvas. The repository keeps large model artifacts outside the application distribution while preserving reproducibility, provenance, and upstream licensing.
BRP Canvas catalog
BRP Canvas reads manifest.json at runtime. Models added to that manifest appear
automatically in Download models, including new categories and families. Each
entry declares its remote files, safe destination below the app's models folder,
expected byte size, and (for weights) SHA-256 digest. Downloads are resumable,
atomically installed, and verified before use.
Important licensing note
There is no blanket license covering every model weight in this repository. Each model directory contains its upstream license and source revision. The root documentation and manifests describe the conversions; the converted weights remain subject to their respective upstream terms.
Available models
All three exports use ONNX opset 20 and return a sigmoid-normalized alpha matte. RGB compositing, resizing, padding, and edge refinement are intentionally handled by the consuming application.
Repository conventions
SeedVR2 3B is provided as a staged FP16 ONNX package: dynamic-spatial VAE
encoder/decoder graphs, input/output projections, and 32 independently
loadable transformer blocks. This lets the native BRP Canvas runtime stream
weights without shipping Python or PyTorch. The conversion is derived from
the Apache-2.0
ByteDance SeedVR2 project.
Each variant contains:
model.onnx and external-data shards when required;conversion.json with source revision, tensor schema, hashes, and conversion provenance;- a model-specific
README.md;
- the unmodified upstream license at the family directory level.
Future shared components, such as the CogVideoX VAE required by LuxDiT, live once under the model family's shared directory and are referenced by the root manifest rather than duplicated.
