Open-Source Fooocus-Style Studio for Qwen-Image 2.1, Built Almost Entirely by Claude
Repulsive_Junket7919 · reddit · 2026-09-27
A developer open-sourced Fooocus-Qwen-Image-2.1, a local Gradio studio inspired by Fooocus — short prompt, one button, complexity under "Advanced." Notably, the entire implementation (code, tests, benchmarks, design research notes) was written by Claude Opus 5.5, with the author only directing and reviewing.
- Text-to-image: three quality presets plus Turbo (6-step distilled LoRA, 10s at 1280×832), 277 Fooocus styles, up to 8 images per run
- Up to 10 reference images addressed as <image1>…<image10>; a 46-pose OpenPose library and sketch canvas; DWPose recognizes poses on your own photos to add library tiles
- Instruction editing in four modes: mask (bytes outside mask untouched), color-coded annotation, exact region crop editing, full-frame; outpainting on any side
- Custom constant-cost mask brush (stock Gradio lagged at 4K); any OpenAI-compatible LLM rewrites prompts; parameters stored in PNG for one-click restore
- Hardware: RTX 3090 24GB + 64GB RAM (pinned text encoder in RAM); INT8 + SageAttention cuts peak VRAM to 10.4 GiB
- MIT code, but Qwen-Image 2.1 weights are under the Qwen Research License — check before commercial use
More from coding & agent
- Sydney Codex Community Buildathon set for Nov 14 with Cerebras-backed ultrafast compute track — paw_lean · 2026-11-11
- shadcn's improve hits 10k stars: let top models plan, cheap models execute code fixes — shadcn · 2026-11-11
- OpenAI open-sources Codex Security CLI and SDK for AI-driven vulnerability scanning and fixing — openai · 2026-11-11
- Claude Opus 5.5 runs an architecture search and finds AttnRes-Canon, a new top Transformer design — SonglinYang4 · 2026-11-11
- 5 research habits worth knowing: read papers backwards, reproduce first, wire coding agents via MCP — cneuralnetwork · 2026-11-11
- 'Adam can now do CAE': AI agent showcase in engineering simulation draws buzz — burhop · 2026-11-11