Gradio open-sources doodle-in LoRA for Qwen-Image 2.1 with full training data
Gradio · x · 2026-09-29
Gradio released Qwen-Image-2.1-doodle-in-LoRA on Hugging Face: a scribble-to-object edit LoRA for Qwen-Image 2.1 that lets you draw an outline on a photo, name the object, and watch it appear in the photo's own art style.
- Fully open: the adapter (diffusers and ComfyUI files), all 6,042 training pairs plus 160 test pairs, every eval output, and the build/training/eval scripts
- Works best with checkpoint 500 + Viggle turbo, finishing edits in 6 steps — image editing control that's hard to get from prompts alone
- Released under the Qwen Research License (non-commercial); training photos from Open Images (CC BY 2.0)
- A live ZeroGPU demo is available: draw a magenta outline, name the object, done
Related event: Gradio Open-Sources Qwen-Image 2.1 Doodle-In LoRA(3 posts)→
More from Multimodal
- Voice model Samuel learns human-like mouth positioning without explicit training — deliprao · 2026-09-29
- 7 ChatGPT prompts that turn ordinary photos into scroll-stopping ad-quality images — aitrendz_xyz · 2026-09-29
- 13 practical ways to use Omni video models for cinematic video creation — CodeByPoonam · 2026-09-29
- Invideo's AI feature film ep. 6: why a DOP still matters when AI builds the world — aziz4ai · 2026-09-29
- Full AI-generated Star Wars anime parody 'The Skywalkers' features real voice actors — JarJarJargon · 2026-09-29
- Midjourney 8.2 realism recipe: imperfect skin texture is the key ingredient — michaelrabone · 2026-09-29