UtopiaIA/Wan2.1-VACE-1.3B-web

Wan 2.1 VACE 1.3B (rCM, 4 steps) for the browser (ONNX)

The VACE module of Wan 2.1 VACE 1.3B by Wan-AI, woven into the few-step rCM transformer of Wan 2.1 T2V 1.3B, converted to ONNX to run with ONNX Runtime Web on WebGPU, as used by Elffuss to make a video between keyframes (first and last frame, or the last frames of the clip before). VACE's 30 base blocks are those of Wan 2.1 T2V 1.3B, so this repository holds only the VACE weights: its graphs read the transformer's weights from UtopiaIA/Wan2.1-T2V-1.3B-web (dit/part0…2.onnx_data, dit/ctx.onnx_data, unchanged), which also has the text encoder and the decoder.

File	What it is
dit/vctx.onnx (+ .onnx_data)	The prompt graph of the 15 VACE blocks: text states in, their cross-attention keys and values out (also reads ctx.onnx_data)
dit/vpart0…2.onnx (+ .onnx_data)	The transformer in three runs with the VACE blocks woven in; each also reads the T2V repository's part<i>.onnx_data. VACE linear layers with 8-bit weights (MatMulNBits)
vae/lighttaew2_1_encoder.onnx	The lighttaew2_1 encoder, 4 frames at a time with its memories: the known frames to latents

New bytes: 813 MB. The sampler is the rCM one of the T2V repository (4 steps, no guidance). Because of the quantised weights, a seed does not give the video of the original model.

Licence

Apache-2.0, as the originals. See NOTICE for the sources and their authors.