Qwen-Image 2.1: open-weights 7B unified gen+edit model, 30s on a 5080 laptop
Lexius2129 · reddit · 2026-09-21
Qwen-Image 2.1 shipped with Day-0 ComfyUI support. Built on February's Qwen-Image 2.0 architecture but released as open weights (research-only license):
- Unified generation and editing in one compact 7B model; up to 10 input images
- Much better textures than prior Qwen-Image-Edit; strong text rendering for infographics, posters, Chinese calligraphy
- Native 2K output and alpha-channel/transparency generation
Benchmarks on a 5080 laptop (16GB VRAM, 32GB RAM) with qwenimage2.1int8convrot: 32s average with the int8 text encoder, 36s with Kijai's w4a8 quantized encoder, 46s with bf16 — and 30s with w4a8 plus the 'Model Attention Backend' node set to comfy kitchen attention.
To try it: update ComfyUI or use Comfy Cloud, grab weights from Comfy-Org/Qwen-Image-2.1 on Hugging Face, and load the text-to-image or image-edit workflow templates.
More from Multimodal
- Redditor benchmarks Minimax H3 Image samplers side-by-side at 16MP+ with identical jaguar prompt — Phantom-log-9484 · 2026-11-12
- Indie Britpop AI music video built with Nano Banana Pro, Kling, and free lip-sync via WAN2GP + LTX 2.5 — cocktailpeanut · 2026-11-12
- P2Echo: one text-prompted model segments multi-view echocardiography with high accuracy and low cost — UBC-V · 2026-11-12
- Redditor makes fast-motion short film with H3 and ships Longshot Studio v0.18.5 with face swap — R34vspec · 2026-11-12
- Reimagining Knight Rider's opening as a Netflix drama with Claude Fable 5.2 — justin_hart · 2026-11-12
- Gemini 4 Carbon builds a full 3D aquarium in Blender and Three.js in ~6 hours — sunjiao123sun_ · 2026-11-12