Wan 2.1 is the state-of-the-art model for generating high-quality videos in minutes. We've taken it a step further, optimizing it to generate high-quality images in seconds. In short, Wan Image is the only model capable of generating 2 megapixel images in 3.4 seconds on a single H100.

From Wan Video to Wan Image…

Wan Video model can generate amazing videos composed of long sequences of image frames. While these image frames are exceptionally high quality since they aim to represent objects behaving consistently according to the rules of physics, each of them requires some computing power. Hence, we decided to turn the Wan Video generating video in minutes into the powerful Wan Image generating images in seconds. To achieve this, we pruned the original model from its video components and applied the best compression methods from the Pruna package, transforming the original video-generation process into an optimized image-generation system that is incredibly fast!

What is Wan Image suitable for?

• Ultra-high-quality images up to 2 Megapixels resolution.

• Consistent, cinematic-quality results, informed by HD video consistent with physics.

Check the quality yourself below! If you want more, you can view a full grid of 100 images generated by Wan Image and Wan Image Juiced on this page.

The Fastest 2 Megapixel Image Generation Model!

In addition to producing high-quality images, Wan Image has been optimized for speed. After compression, it generates a 1 Mgp image in 1.7 seconds and a 2 Mgp image in 3.4 seconds using just one H100. This is 3.6x faster than Seedream and 1.41x faster than Flux-1.1 Pro for 1 Mgp images, and 2.1x faster than SeedDream and 1.6x faster than Flux-1.1 Pro for 2 Mgp images.

We have released this model with a competitive pricing model through Replicate. Generate a 1K image for $0.02 and a 2K image for $0.025, making it the most cost-efficient option for 2K image generation.

We conducted a quantitative benchmark comparing Seedream, FLUX 1.1 Pro, Wan Image, and Wan Image Juiced alongside qualitative image assessments. The benchmark supports 2 megapixel generation based on the first 100 prompts of the GenAI-Bench dataset. We used every model with its default settings to generate 1440x1440 images with a fixed seed of 0. The results highlight Wan Image and Wan Image Juiced as generating state-of-the-art images, while being one order of magnitude more efficient.

For each combination of model and metric, we show its scores and ranking in parentheses.

Disclaimers

Although we are releasing this practical research implementation with good intentions, we do want to emphasise some disclaimers about the research tricks we implemented to empower this release.

• The original works: Our Wan Image implementation is based on the original work of the Wan-Video team. All original limitations and licensing terms from the base model continue to apply to this adaptation.

• Prompting considerations: We haven't conducted an extensive evaluation of how video-optimized prompts translate to static image generation. The original model was designed for video, so the output resulting from static prompts might not be fully represented by the selected frame of a video model.

• Try it in your specific use case: We recommend testing it thoroughly, validating your outputs, and reporting any unexpected behaviours or limitations to us.

In short, have fun, make the most of the model, and help us improve it when needed! Share your findings and use cases to advance the Wan Image model applications on social media, too, and we would be happy to reshare!

Enjoy the Quality and Efficiency!

Want to take it further?

• Try our Replicate endpoint with just one click.

• Compress your own models with the Pruna package.

• Stay up to date with the latest AI efficiency research on our blog, explore our materials collection, or dive into our courses.

Do you think we are rocking it? Please give us a star to show your support!