Walkyrie-1.3B-v1.0(Preview)Text-to-Image
A 1.3B parameter model turns video AI into an image generator — but it's only 20% trained.
Walkyrie-1.3B is a new text-to-image diffusion model built by repurposing an existing text-to-video architecture. The developer, kpsss34, started with Wan2.1-T2V-1.3B — a compact video generation model — and pruned its UMT5 text encoder down to roughly 1 billion parameters. The entire pipeline was then retrained to produce static images instead of video, effectively converting a temporal generative model into a spatial one. This approach leverages existing video model knowledge while reducing computational overhead.
The model is currently an early preview, trained to only about 20% of its intended training budget. As a result, quality and stability are not yet production-ready. The developer openly notes that anatomy — a classic challenge for small-scale image models — remains the biggest weakness. The release aims to gather community feedback and encourage further development. Despite its limitations, Walkyrie-1.3B demonstrates an efficient strategy for adapting video diffusion models to image generation, potentially opening new paths for compact, multi-purpose generative systems.
- Derived from Wan2.1-T2V-1.3B, a compact text-to-video diffusion model, with UMT5 encoder pruned to ~1B parameters.
- Currently trained to only 20% of the planned budget — an early preview for community testing and feedback.
- Main limitation is anatomy generation, a common issue for small-scale image models; quality expected to improve with further training.
Why It Matters
Shows how video diffusion models can be repurposed for image generation, but underscores the need for full training to achieve reliable output.