Kimodo-SOMA-RP-v1.1-ONNX

Kimodo-SOMA-RP-v1.1 ONNX

KitsuMate conversion of NVIDIA Kimodo-SOMA-RP-v1.1, checkpoint revision 6c9233af1180b8151e3c4703477104af5dce9dd5. NVIDIA and the Kimodo authors retain authorship. No model retraining was performed.

Variable-length denoiser

• onnx/model_fp32.onnx: FP32 reference.
• onnx/model_fp16.onnx: FP16 deployment.

Both replace the former fixed-60-frame exports. Batch size is fixed at three (text, constraint, unconditional CFG branches). Temporal length T is dynamic and validated for 2–300 frames at 30 FPS. Longer animations require conditioned window sequencing; increasing T beyond the trained window is not supported.

The graph exports one denoiser evaluation. CFG, DDIM scheduling, constraint compilation, continuation, motion decoding and contact cleanup are host responsibilities. The existing LLM2Vec encoder remains compatible. Unused text tokens are zero-padded; the first branch receives the text embedding, the second the constraint mask, and the third is unconditional.

Export and verification

Official source: nv-tlabs/kimodo revision 1aece8c124d73d255ceff5086d983b844c9f4e94. PyTorch 2.5.1+cu124, ONNX 1.16.1, opset 17, ONNX Runtime GPU 1.24.4, RTX 3070 Laptop GPU. Export-only attention uses explicit Q/K/V projections and dynamic reshapes, with the original trained parameters, to avoid fixed internal sequence lengths in the legacy exporter. Numerical references use the unmodified NVIDIA attention implementation.

Lengths 2, 30, 59, 60, 61, 150, 299 and 300 were verified, including repeated length changes in one session. Inputs include constraint masks and a nonzero initial heading. All outputs were finite with the correct shape.

FP32 single-step: minimum cosine 1.00000000; maximum MAE 0.00000404.

FP16 single-step: minimum cosine 0.99988377; maximum MAE 0.00324112.

Full 25-step separated-CFG/DDIM sampling was compared with NVIDIA Python at 60 and 300 frames using identical initial noise, a prebaked text embedding and root constraints:

Timings exclude session creation, text encoding and postprocessing and are measurements on the listed laptop GPU, not real-time guarantees. FP16 diffusion trajectories can differ from FP32.

SHA-256

• model_fp16.onnx: c4a4aee64a6e7a2d1551d6382d7cd89e35f3d5e0e50b6c62728bc5563b839497
• model_fp32.onnx: 07341f1d89bea7a2a392a63040f39c048f8b0638ee1f50a5a0988bb8a9d7109b

Q8 weight-only variant

onnx/model_int8.onnx is an optional size-focused variant: 305,933,441 bytes, 47.9% smaller than FP16. It uses per-output-channel symmetric INT8 weights for 140 linear tensors, standard ONNX opset 17 dequantization, and FP16 computation. Inputs and outputs retain the FP16 model interface and dynamic temporal length; it does not extend the trained window.

Basic final-output comparisons against FP16 used the same text embedding, noise and constraints, with 25 denoising steps. On CUDA, cosine similarity was 0.994910 / 0.998266 and mean absolute error was 0.045806 / 0.024012 at 60 / 300 frames. DirectML cosine similarity was 0.995351 / 0.998220, with mean absolute error 0.043532 / 0.024482.

On an RTX 3070 Laptop GPU, median of three warm runs after one full warmup:

Timings exclude loading, text encoding and pose cleanup. Compare variants within each backend; host implementations differ. These checks used one prompt/seed per length, not broad perceptual validation. Mobile performance and memory use have not been measured. Smaller weight storage does not guarantee less runtime memory or faster inference.

SHA-256 (model_int8.onnx): 4c59427c7d9a33feddd50a184b806ff135b53408c1772ccd0f2a8fd1f8a44930.

License

Distribution and use are subject to the included NVIDIA Open Model License and NVIDIA Trustworthy AI terms referenced by that license.