Pruna AI turns Wan 2.1 into fast image model
Pruna AI has stripped the video parts out of Wan 2.1 and compressed what remained into an image generator it says makes 2-megapixel images in 3.4 seconds on one H100.
Pruna AI announced Wan Image Juiced, an image-generation model built from Alibaba's Wan 2.1 video model. The company says it removed the model's video components and then applied compression methods from its own Pruna package. This makes Wan Image a derivative of an existing model, not a new one trained from scratch.
According to Pruna, the compressed model generates a 1-megapixel image in 1.7 seconds and a 2-megapixel image in 3.4 seconds on a single H100. The company says that makes it 3.6x faster than Seedream and 1.41x faster than FLUX 1.1 Pro at 1 megapixel. At 2 megapixels, it claims speedups of 2.1x over Seedream and 1.6x over FLUX 1.1 Pro. The model is available on Replicate at $0.02 per 1K image and $0.025 per 2K image.
Pruna ran its own quality benchmark on the first 100 prompts of GenAI-Bench, generating 1440x1440 images with default settings and a fixed seed. In those results, Seedream ranked first on VQA, CLIP and ImageReward. Wan Image Juiced placed second on VQA, ARNIQA, CLIP IQA and ImageReward. The uncompressed Wan Image led on ARNIQA and CLIP IQA. The results do not show Wan Image Juiced beating Seedream on overall quality. Its main pitch is speed and cost.
Pruna says the original Wan licensing terms and limitations still apply. It also says it has not thoroughly tested how prompts written for video carry over to still images.