Method
CoFi separates long-horizon composition into global structure formation and local detail recovery.
Generated scaffold: globally coherent but locally blurred
Stage 1: Coarse Scaffold Construction
CoFi denoises all local plans in parallel and pulls their clean estimates toward a shared scaffold. This stage resolves long-range structure first, producing a coarse but globally aligned plan.
Final output: globally coherent with fine local detail
Stage 2: Structure-Preserving Refinement
CoFi re-noises the coarse scaffold to an intermediate timestep and denoises it again with the same pretrained local prior. This second pass restores local detail while keeping the global arrangement fixed.
NFE–Performance Comparison
We compare performance against the number of function evaluations (NFE) across the three domains. CoFi improves the main coherence metrics across domains while using 2–8× fewer NFE than CDGS, substantially reducing inference cost. The scaffold construction stage uses the same denoiser evaluations as GSC, and the refinement stage adds only t* extra steps, giving a total cost of T + t*.
Left: Robotic planning — CoFi achieves 96% success with 2.1× fewer NFE than CDGS. Center: Panoramic images — CoFi reduces Intra-LPIPS to 0.45 with 8× fewer NFE. Right: Long videos — CoFi reaches 94.1% subject consistency with 8× fewer NFE.
Faster and more coherent
Side-by-side comparison of 273-frame long video generation. CoFi better preserves subject appearance and scene layout over long temporal horizons while using 8× fewer denoiser evaluations.
"A cute happy panda, dressed in a small red jacket and a tiny hat, sits on a wooden stool in a serene bamboo forest, strumming a miniature acoustic guitar..."