Local ComfyUI workflow: Qwen 2.1 nails product references but fails on people
Busy_Site_8767 · reddit · 2026-10-08
A creator making product lifestyle photos for a cosmetics brand shares a fully local ComfyUI pipeline on an RTX 5080 16GB.
Current workflow
- Edit real reference photos with Qwen Image 2.1 (scene + product + color reference) — best-in-class at reproducing product shape, labels and materials
- Upscale with SeedVR2, add labels afterwards
Pain points
- People are the weak spot: broken legs/hands/fingers, airbrushed skin
- Qwen 2.1's license doesn't allow commercial use without a separate agreement
Options under consideration
- Build scenes (with or without people) in Krea 2 first, then insert products via Qwen 2.1 reference editing
- Or a low-denoise Krea 2 img2img pass over finished Qwen images to fix skin and anatomy
- Also seeking recommendations for non-explicit intimate scenes (silhouettes behind frosted glass, sleeping couples) and LoRAs for realistic skin and hands
More from Multimodal
- SheetSage2 uses synthetic supervision to lead music transcription on 12 of 15 benchmarks — m-a-p · 2026-10-08
- Marigold V2 Rethinks DiT for Depth Estimation, 16-26% AbsRel Gains on KITTI/ETH3D — Cohere_Labs · 2026-10-08
- Qwen 2.1 image experiment: multi-angle character sheets with a distinctly donghua style — Gkashhh · 2026-10-08
- Nano Banana 2.1 image generation available for free in Google Flow — aziz4ai · 2026-10-08
- Local LLM + Tripo3D: A Mini 4WD-Style 3D Racing Game Built in One Day — lxfater · 2026-10-08
- ChatGPT finds 10 anatomical failures in a GPT Image 2.5 generation — cocosus · 2026-10-08