KV Cache Quantization for Self-Forcing Video Generation: A 33-Method Empirical Study

TL;DR

This paper empirically evaluates 33 quantization and cache policies for KV-cache compression in self-forcing video generation, aiming to improve memory efficiency without sacrificing quality.

Contribution

It provides a comprehensive empirical analysis of KV-cache compression methods, identifying practical approaches like FlowCache-inspired soft-prune INT4 for better memory and performance balance.

Findings

FlowCache-inspired soft-prune INT4 achieves 5.42-5.49x compression with reduced VRAM.

High-fidelity methods like PRQ_INT4 and QUAROT_KV_INT4 are costly in runtime or memory.

Nominal compression methods still exceed peak VRAM due to current integration practices.

Abstract

Self-forcing video generation extends a short-horizon video model to longer rollouts by repeatedly feeding generated content back in as context. This scaling path immediately exposes a systems bottleneck: the key-value (KV) cache grows with rollout length, so longer videos require not only better generation quality but also substantially better memory behavior. We present a comprehensive empirical study of KV-cache compression for self-forcing video generation on a Wan2.1-based Self-Forcing stack. Our study covers 33 quantization and cache-policy variants, 610 prompt-level observations, and 63 benchmark-level summaries across two evaluation settings: MovieGen for single-shot 10-second generation and StoryEval for longer narrative-style stability. We jointly evaluate peak VRAM, runtime, realized compression ratio, VBench imaging quality, BF16-referenced fidelity (SSIM, LPIPS, PSNR), and terminal drift. Three findings are robust. First, the strongest practical operating region is a FlowCache-inspired soft-prune INT4 adaptation, which reaches 5.42-5.49x compression while reducing peak VRAM from 19.28 GB to about 11.7 GB with only modest runtime overhead. Second, the highest-fidelity compressed methods, especially PRQ_INT4 and QUAROT_KV_INT4, are not the best deployment choices because they preserve quality at severe runtime or memory cost. Third, nominal compression alone is not sufficient: several methods shrink KV storage but still exceed BF16 peak VRAM because the current integration reconstructs or retains large BF16 buffers during attention and refresh stages. The result is a benchmark harness, analysis workflow, and empirical map of which KV-cache ideas are practical today and which are promising research directions for better memory integration. Code, data products, and the presentation dashboard are available at https://github.com/suraj-ranganath/kv-quant-longhorizon/.