Gradio Open-Sources Doodle-in LoRA for Qwen-Image 2.1
Gradio's team released the Qwen-Image-2.1-doodle-in-LoRA on Hugging Face, enabling doodle-based image editing: draw a rough magenta outline on a photo with an object name, and the model replaces that region with the named object following the doodle's shape and pose while preserving the original lighting, scene, and art style.
Confirmed
- The editing workflow takes only 6 steps with the accompanying tools
- The LoRA's data and training scripts are fully open on Hugging Face
- A live demo is available on Hugging Face ZeroGPU
- According to a repost, the project was allegedly designed, trained, and evaluated autonomously by a "single-prompt-driven ML Intern" (from a secondhand account)
Why it matters
This work pushes generative editing control granularity from text prompts down to hand-drawn doodles, giving creators a more intuitive way to do local replacement; fully open data and scripts lower the barrier for reproduction and secondary development, and if the autonomous-ML-intern claim holds, it also hints at AI agents automating ML workflows.
2026-09-29 ~ 2026-09-30 · 5 related posts
Primary sources
- [source] Doodle In LoRA for Qwen-Image 2.1 edits photos in 6 steps while keeping art style — Gradio · 2026-09-29
- [source] Gradio open-sources doodle-in LoRA for Qwen-Image 2.1 with full training data — Gradio · 2026-09-29
- [source] Try the Qwen-Image 2.1 doodle-in LoRA live on Hugging Face ZeroGPU — Gradio · 2026-09-29
- One-prompt-trained Qwen-Image-2.1 LoRA turns your doodles into photorealistic objects — Gradio · 2026-09-30
- Doodle-in LoRA for Qwen-Image-2.1 turns scribbles into photorealistic objects — Gradio · 2026-09-30