AccVideo: Accelerating Video Diffusion with Synthetic Dataset
AccVideo by aejion accelerates Wan2.1 video generation by up to 8x using a synthetic dataset approach. Available in 14B fp8 with rank32 LoRAs for faster inference.
AccVideo
Wan2.1Video GenerationAccelerationFast InferenceOpen SourceAccelerated video generation model by aejion that speeds up Wan2.1 inference by up to 8x. Uses a synthetic dataset distillation approach to reduce sampling steps while preserving visual quality. Available in 14B fp8 with rank32 LoRAs.
| Developer | aejion |
| Architecture | Wan2.1-based with synthetic dataset acceleration |
| Model Sizes | 14B (fp8), plus LoRAs (rank32) |
| Capabilities | 8x accelerated text-to-video generation |
Overview
AccVideo (Accelerating Video Diffusion Model with Synthetic Dataset) is an optimization for Wan2.1 that dramatically reduces inference time. By training on a carefully curated synthetic dataset, the model learns to produce high-quality video output with significantly fewer sampling steps. The result is up to 8x faster video generation compared to the standard Wan2.1 pipeline, making it practical for iterative creative workflows and real-time applications.
ComfyUI Integration
AccVideo is available for ComfyUI through the Kijai WanVideoWrapper. The optimized fp8 checkpoint combined with rank32 LoRA adapters provides an excellent balance of speed and quality.
Available Variants
| Variant | Precision | Description |
|---|---|---|
| 14B Base | fp8 | Accelerated core model |
| LoRAs | rank32 | Speed-optimized fine-tuning adapters |
Resources
| Resource | Link |
|---|---|
| GitHub Repository | github.com/aejion/AccVideo |
| Hugging Face (Original) | aejion/AccVideo-WanX-T2V-14B |
| Hugging Face (Kijai) | Kijai/WanVideo_comfy |
Guides and workflows related to this model series.