MiniMax-H3-VAE-ONNX
ONNX version of the MiniMax-H3 VAE for ComfyUI, which can increase speed by up to 1.7x
Usages
- Put encoder & decoder
.onnxfiles (and.datafiles, if present) intoComfyUI/models/vae
You can use the
w4a16_awqdecoder if you don't have 12GB+ of VRAM.
- Install this extension: ComfyUI-H3VAE_TRT
Benchmark
Tested at 1344×768 @ 5s empty latent:
| Model Name | Size | Dec Time | Speedup | Rel RMS Error | PSNR (dB) | Has NaN |
|---|---|---|---|---|---|---|
| fp16_baseline | 4.85 GB | 17.87s | 1.00x | 0.000% | Inf (Exact) | False |
| trt_fp16_engine | 4.85 GB | 11.84s | 1.50x | 0.117% | 65.84 dB | False |
| trt_w4a16_engine | 1.54 GB | 13.10s | 1.36x | 6.778% | 30.65 dB | False |
License
The weights in this repository are released under the Apache 2.0 License. Use of the MiniMax-H3 base model is also subject to its corresponding license and terms of use.