UtopiaIA/Wan2.1-T2V-1.3B-web

Wan 2.1 T2V 1.3B (rCM, 4 steps) for the browser (ONNX)

Wan 2.1 T2V 1.3B by Wan-AI, with the few-step rCM weights for its transformer and the lighttaew2_1 tiny decoder, converted to ONNX to run with ONNX Runtime Web on WebGPU, as used by Elffuss. The tokenizer is read from the original Wan repository at revision 37ec512624d61f7aa208f7ea8140a131f93afc9a; this repository holds only the converted weights.

File	What it is
t5/encoder_0…3.onnx (+ .onnx_data)	The UMT5-XXL text encoder in four shards of its layers, linear layers with 4-bit weights (MatMulNBits)
t5/embed_q4.bin	The UMT5 token embedding table in 4-bit blocks, read by the page
dit/ctx.onnx	The transformer's prompt graph: text states in, the cross-attention context out
dit/part0…2.onnx	The transformer blocks in three runs, linear layers with 8-bit weights (MatMulNBits)
vae/taew2_1_light.onnx	The lighttaew2_1 decoder, one latent frame at a time with its memories

The sampler follows the rCM / TurboDiffusion inference script for Wan 2.1 (2 to 4 steps). Because of the 4-bit text encoder and the 8-bit transformer, the video for a given seed is not the one of the original model, though it follows the prompt as well.

Licence

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