Kimodo-SOMA-RP-v1.1 ONNX
KitsuMate ONNX export of nvidia/Kimodo-SOMA-RP-v1.1, pinned at 6c9233af1180b8151e3c4703477104af5dce9dd5.
NVIDIA and the Kimodo authors retain authorship of the model. KitsuMate only provides the ONNX conversion in the standard flat onnx/ layout for the fixed 60-frame runtime.
Variants
onnx/model_fp32_b3_t60.onnx: full FP32 reference denoiser.onnx/model_fp16_b3_t60.onnx: FP16 deployment denoiser.
Both graphs use the three-sample classifier-free-guidance batch and 60-frame
sequence expected by the KitsuMate Kimodo runtime. They were exported from the
pinned NVIDIA checkpoint with the same fixed-shape exporter. The FP32 graph was
validated with ONNX Runtime 1.24.4 CUDA against the PyTorch checkpoint: cosine
similarity 0.9999985, mean absolute error 0.000494, and finite output.
Distribution and use are subject to the included NVIDIA Open Model License and NVIDIA Trustworthy AI terms referenced by that license.