Updated ComfyUI-SeedVR2-VideoUpscaler-with-TensorRT v1.6.6 - Native ConvRot W4A8 Support & DisTorch2 Scope
The other day, when I published information about v1.5.7, I received a request in the comments regarding support for w4a8.
Proof1 community thread
Key points
- Firstly, as for the advantages of w4a8, benchmark results clearly show that the quality is superior to that of NVFP4.
- On the other hand, what wasn’t as good as I’d hoped was that VRAM consumption was almost on a par with ConvRot INT8 (though, given that w4a8 expands weights to 8 bits, this was to be expected), and that with Distorch2—which I’d gone to such lengths to implement—VRAM usage actually skyrocketed, rendering it unusable.
- The on/off function for Norm BF16, which was previously implemented in Distorch2 to counteract VRAM spikes, has now also been implemented in the Legacy Loader.
- Enabling this feature allows you to suppress VRAM spikes by up to approximately 3GB, albeit at the cost of a slight reduction in image quality.
Sources (1)
- [1]Updated ComfyUI-SeedVR2-VideoUpscaler-with-TensorRT v1.6.6 - Native ConvRot W4A8 Support & DisTorch2 Scoper/StableDiffusion (top, day) · Oct 11, 01:14 PM
The other day, when I published information about v1.5.7, I received a request in the comments regarding support for w4a8.
Firstly, as for the advantages of w4a8, benchmark results clearly show that the quality is superior to that of NVFP4.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 11, 2026Vision-Predictor-Action: Towards Efficient Embodied AI via Neuro-Symbolic JEPA Predictors
- Oct 10, 2026Towards Privacy-Preserving Thermal Human Perception from Dataset to Deployment
- Oct 10, 2026LongHisDoc: A Comprehensive Benchmark for Chinese Long Historical Document Understanding
- Oct 10, 2026EviScene: Beyond "Is It AI?" for Agentic Image Factuality Assessment
- Oct 9, 2026Qwen/Qwen-Image-2.1-Turbo
- Oct 9, 2026AMLworld-Compact: Importance-Weighted Downsampling for Cost-Effective LLM Evaluation and Error Diagnosis
