AnalysisOpinion / analysisImage, Video & 3D Generation · Efficiency & Inference · Hardware & Compute1 source · Oct 11, 2026

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)

Extractive summary: sentences quoted from the sources.

Before this

  1. Oct 10, 2026Krea2 Turbo Distill 2 step LoRA - FINAL checkpoint released (chk51195)
  2. Oct 9, 2026Qwen/Qwen-Image-2.1-Turbo
  3. Oct 8, 2026unslothai/unsloth v0.1.905-beta: Sandboxing is here!
  4. Oct 8, 2026ConwayResearch/Underdog-Saluki-27B-1.0
  5. Oct 7, 2026[AINews] Claude Haiku 5.5 — better than GPT-6 Luna at the same pricing
  6. Oct 7, 2026unslothai/unsloth v0.1.904-beta: Train your own Decision model

Related