SYLIN

ComfyUI

by Comfy Org · Runtime · Image

promising

Widely used way to run image models locally; its new speed claims are untested

Rewrote LLM text-generation hot path for Qwen3.x; same checkpoints generate up to 2x faster on the same GPU, same quality

Why we say so

AI Search and Fahd Mirza both got Qwen Image, MiniMax H3 Image and MOSS Image running through ComfyUI templates or workflows after updating. The 0.45.0 claims of up to 2x faster text generation (Qwen3.5 9B from 106 to 219 tok/s on an RTX 5090) come only from Comfy Org, with no independent check yet.

Evidence

  1. Nov 15
    MiniMax-H3-Image runs via ComfyUI by downloading model, VAE and text files into the models folders plus a provided workflow
  2. Nov 11
    New W4A8 GEMV tensor-core kernel with interleaved weights and scales, fused RMSNorm/SwiGLU prologues; M=8 verify rows cost about as much as M=1
  3. Oct 20
    Ships MOSS Image text-to-image and layout-guided workflow templates after updating; layout node needs Nodes 2.0 enabled
  4. Sep 22
    Qwen Image 2.1 ships with ComfyUI templates for text-to-image, edit and background removal
  5. Aug 23
    ComfyUI open-sourced Comfy MCP, letting agents like Codex see local workflows, models, nodes and GPU and build workflows from prompts
Show all 6
  1. Aug 5
    Updated ComfyUI includes MiniMax H3 text-, image- and reference-to-video templates alongside cloud API versions

Runs on ComfyUI

5 models · newest first

Alternatives

QwenIn ComfyUI, Qwen3.5 9B rose from 106 to 219 tok/s on RTX 5090 and 91 to 129 tok/s on RTX 5080 (v0.40.0 to v0.45.0)