Model Reveals What to Cache: Profiling-Based Feature Reuse for Video Diffusion Models

Ser-Nam Lim, Zihao Wang, Yexin Liu, Harry Yang, Mingzhe Zheng, Xianfeng Wu, Xuran Ma, Yaofu Liu, Everlyn Ai

Code that ran

We lifted 1 functions out of this paper's own repositories and ran 1 of them in a sandbox. "Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results.

Repository	Role	Ran
geekguru123/profilingdit	—	1 of 1

Function	Status	Where it lives
DitCache	Ran	geekguru123/profilingdit/HunyuanVideo/hyvideo/ditcache.py pointer only (licence: NONE) · get_code("18b501ee2094dc9a")

Repositories linked to this paper

geekguru123/profilingdit — marked official in Papers with Code; named in the paper

Some links come from the archived Papers with Code dataset (CC BY-SA 4.0): attribution and licence.

Abstract

Figure 1. Comparison of visual result quality across different methods. The video has a resolution of 780p and consists of 129 frames, with one representative frame extracted from each video for visualization. Our method consistently outperforms TeaCache [15] in both visual quality and efficiency. Latency is evaluated using a single H200 GPU. All results are generated with seed 42.

For agents

The same record, over MCP at https://syntology.ai/mcp:

get_harvested_code_for_paper("2504.03140")
get_code_for_paper("2504.03140")
have("2504.03140")

Connect an agent — have() is free.