Expired50% confidenceFactExact time
ComfyUI-INT4-Fast was independently improved upon BobJohnson24's ComfyUI-INT8-Fast, and relies on comfy-kitchen under the hood to invoke the GPU's Tensor Cores.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
7/12/2026
First Seen
Valid until: 10/10/2026(expired)
Sources
ComfyUI-INT4-Fast：6GB显卡流畅运行Flux大模型实测
redditr/StableDiffusion7/10/2026
Related Claims
UnverifiedComfyUI-INT4-Fast supports native mixed-precision checkpoints, determining the compute path at per-layer granularity: backbone modules are routed to INT4 Tensor Cores, while sensitive layers (patch projection) are directed to Triton's INT8 execution path in INT8 format.75% similarUnverifiedComfyUI-INT4-Fast is a custom node package that brings INT4 (W4A4) quantized inference into the ComfyUI ecosystem, enabling even a 6GB VRAM RTX 3060 to run Flux-based Krea2 Turbo models.69% similarUnverifiedThe 16-series Turing architecture introduced first-generation Tensor Cores supporting FP16 mixed-precision training, while the 30-series Ampere architecture upgraded to third-generation Tensor Cores and added hardware acceleration support for BF16 and TF32.69% similarUnverifiedThe NVIDIA Ada Lovelace architecture (RTX 40 series) introduced dedicated INT4 Tensor Cores on consumer-grade GPUs for the first time, theoretically achieving up to 4× the matrix operation throughput compared to FP16.67% similarUnverifiedNVIDIA designed a dedicated Tensor Core instruction mma.kind::mxf4 for NVFP4 in the Blackwell architecture (SM 100/103, such as the B100 and B200 series), and adopted an FP8 scaling factor mechanism with one scale per group of 16 elements.67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/489322API
curl https://kongchang.com/api/v1/knowledge/claims/489322MCP
get_claim(id=489322)