Spatiotemporal Skip Guidance

STG (CVPR 2025): training-free guidance for video diffusion transformers — self-perturbs the model by skipping spatiotemporal layers to create an implicit weak model, avoiding CFG's over-saturation while boosting fidelity and dynamics. Real adoption: upstreamed into HuggingFace Diffusers, Lightricks LTX-Video, and ComfyUI. From Jaegul Choo's DAVIAN Lab.