SkyReels-V1
Project lead · 2024 - 2025
Open-source human-centric video foundation model, released with text-to-video and image-to-video weights and an inference framework.
- Video diffusion
- DiT
- Post-training
- Inference optimization
SkyReels-V1 is a human-centric video foundation model, built by fine-tuning HunyuanVideo on roughly ten million film and television clips. Released in February 2025 with both text-to-video and image-to-video weights.
The work that mattered was upstream of the model. We built a data cleaning and annotation pipeline that classifies facial expressions into 33 types, reconstructs 3D human pose to reason about where people stand relative to each other, and defines over 400 action semantic units. Human-centric video fails in ways that are obvious to viewers and invisible to generic metrics, so the labels have to encode what a viewer actually notices.
Training ran in three stages: domain transfer from general video to human-centric footage, conversion of the text-to-video model into an image-to-video model, then high-quality fine-tuning on a curated subset.
On VBench the text-to-video model scored 82.43 overall, ahead of the open-source models we compared against at release, with its largest margins on dynamic degree and multiple objects.
We also open-sourced SkyReelsInfer, the inference framework. It supports context, CFG, and VAE parallelism, and runs on consumer hardware through FP8 weight-only quantization and parameter-level offload. On four RTX 4090s it cut end-to-end latency by 58.3% against HunyuanVideo with xDiT. Shipping the inference path alongside the weights was deliberate: a model nobody can afford to run is not really open.