One step.
A thousand
possibilities.
From a seed to an image in a single forward pass. Explore W-Flow’s published examples and the included class-conditional generation demo.
01 / Explore the checkpoints
One subject.
Three model sizes.
The same golden retriever class and seed, sampled with each checkpoint’s recommended guidance. Select a model to view its published example.
Model size- Subject
- Golden retriever
- ImageNet class
- 207
- Seed
- 1
- Guidance
- 1.19
Includes searchable class selection, seed and guidance controls, and batches of up to four images.
Sample images from BiliSakura/W-Flow-diffusers. These are provided examples, not a live generation.
02 / A different kind of generation
A class, a seed, a single step.
W-Flow generates images conditioned on ImageNet’s 1,000 classes. The Gradio app lets you search for a subject, choose B, L or XL, and reproduce a result with its seed. Resolution is fixed at 256 × 256.
03 / Built for the model
Explore the research.
W-Flow is introduced in One-Step Generative Modeling via Wasserstein Gradient Flows. This demo uses the published Diffusers checkpoints and their custom pipeline. The model weights and code are released under Apache 2.0.
Run the included Gradio app
The application source and dependencies are included in this Space. Once GPU access is available, use README.gradio.md as the Space’s README.md to activate the Gradio configuration, and select ZeroGPU hardware in Space settings. You can also duplicate this Space to an account with GPU access.