Qwen Draw2Photo: Turn Your Sketches into Realistic Photos
What if you could doodle a rough sketch, snap a photo of someone, and have AI merge them into a realistic image that actually looks like that person? That’s exactly what Qwen Draw2Photo does.
What is Draw2Photo?
Draw2Photo is a new image generation technique from Alibaba’s Tongyi Lab. It takes two inputs:
- A simple sketch — a rough line drawing, stick figure, or doodle
- A reference photo — a photo of a person
The model then generates a realistic image that follows the sketch’s pose and composition — while preserving the identity of the person in the reference photo.
How It Works
The key innovation is identity-preserving LoRA fusion:
- Sketch conditioning — The model understands the spatial layout, pose, and shape from the sketch
- Identity preservation — LoRA adapters extract facial features and appearance from the reference photo
- Fusion — Both signals are combined during generation, so the output matches the sketch’s structure while wearing the reference person’s likeness
Unlike traditional image-to-image translation that only transfers style, Draw2Photo actually understands both inputs and creates a coherent fusion.
Try It Yourself
The team shared two resources:
- Interactive demo: Try Draw2Photo online — upload your sketch and photo, see the result
- LoRA weights: Download the LoRA for local experimentation
Both are publicly available, making this one of the most accessible identity-preserving sketch-to-photo tools released so far.
Why This Matters
- Character design — Game artists can sketch a character, reference a real face, and instantly see a realistic render
- Concept art — Quickly visualize ideas without full rendering pipelines
- Content creation — Generate consistent character visuals across different poses and scenes
- Accessibility — No need for expensive GPUs or complex workflows; the demo runs online
The Bigger Picture
Draw2Photo adds to a growing ecosystem of AI tools that blur the line between intention and output. The ability to combine structural control (sketch) with identity preservation (reference photo) means creators can iterate faster — sketching an idea and seeing a realistic preview in seconds.
It’s another reminder that 2026 is the year AI image generation moves from “generate anything generic” to “generate exactly what I mean.”