Open Sourcing my 10M model for video interpolations with comfy nodes. (FrameFusion)
Solo developer open-sources commercial-grade 37MB model that runs smoothly on consumer GPUs.
Independent developer Hugo Zanini has open-sourced FrameFusion, the 10-million parameter video interpolation model powering his commercial application. The model, which weighs in at just 37MB in fp32 format, represents years of development focused on achieving a practical balance between quality and speed for consumer hardware. Trained entirely on Kaggle using a mix of live-action and anime footage, FrameFusion uses a motion flow architecture that warps original images rather than generating RGB pixels directly—a design choice that reduces computational load and minimizes visual artifacts.
Zanini's release includes not only the model weights but also custom nodes for ComfyUI, making it immediately usable in popular AI workflow tools. The developer acknowledges the trade-off: while the model's small size (under 10M parameters) enables it to run efficiently on low-end GPUs, it may not match the absolute quality of larger, more resource-intensive alternatives. This release follows Zanini's six-year journey in machine learning, which began with the Rife-App collaboration and has now culminated in making professional-grade interpolation technology freely available to the community.
- 10-million parameter model with 37MB file size optimized for low-end hardware
- Trained on mixed live-action and anime datasets using Kaggle's infrastructure
- Includes ComfyUI workflow nodes for immediate integration into existing pipelines
Why It Matters
Democratizes professional video interpolation by making it runnable on consumer hardware instead of requiring expensive cloud GPUs.