Open-source MingImage-01 runs in ComfyUI via Kijai's PR: ~50s for 2048×2048 on a 4090
GreyScope · reddit · 2026-09-24
The newly open-sourced MingImage-01 (Design family) image model now works in ComfyUI via Kijai's PR (#16482, not yet merged). The author ran a fresh Comfy install with the PR, copied the workflow from the PR page, and used a ChatGPT-generated Llamafarm prompt as a minimal proof of concept.
Real-world numbers: 19.7GB VRAM on a 4090, 50 seconds to generate a 2048×2048 image. The workflow used the Qwen38BFluxKlein clip model, though the author suspects a stronger text encoder noted in the PR might be needed.
Related event: Meitu's Open-Source Ming Image Model Now Runs Natively in ComfyUI(3 posts)→
More from Multimodal
- Pooling-aware training cuts retrieval vectors 4x while boosting nDCG to 97.8% of unpooled — lateinteraction · 2026-11-11
- Generating accurate 3D building models from 150 planning PDFs with new Gemini model — fofrAI · 2026-11-11
- Stanford's DyCraft Uses Agentic Decomposition to Cut Physical Errors in Long-Horizon 4D Scene Generation — StanfordUniversity · 2026-11-11
- Midjourney SREF pick: --2523103199 for loose pencil-sketch portraits — tisch_eins · 2026-11-11
- H3 Image Tips: 16MP via ComfyUI Workflow Plus Detailed Prompts Yields Best Results — gatortux · 2026-11-11
- Gemini-generated endless loop of quantum cryostats self-assembling — hapticdata · 2026-11-11