Hugging Face Releases DeCoPrune: Revolutionizing Autoregressive Video Generation with Efficient KV-Cache Pruning

Summary:Hugging Face introduces DeCoPrune, a novel method addressing the growing KV cache issue in autoregressive video generation. By treating cache compression as a denoising-consistency optimization problem, DeCoPrune uses denoising difficulty as a proxy for token retention value, significantly reducing historical KV tokens while maintaining near-full cache long-range recall. Experiments show that DeCoPrune achieves over 85% pruning of historical KV tokens and accelerates continuation generation by m

---

Core Breakthrough

Autoregressive video generation offers significant advantages in streaming generation and interactive control, but its KV cache grows continuously with the generated history, leading to high inference costs. Existing compression strategies either discard history using fixed windows or select tokens through local attention and similarity signals, which fail to directly measure whether the current token contributes information beyond the retained context.

Hugging Face's DeCoPrune addresses this issue through the following:

• Denoising-Consistency Optimization: Treating cache compression as a denoising-consistency optimization problem, DeCoPrune uses denoising difficulty as a proxy for token retention value.
• Dynamic Pruning Mechanism: By measuring the discrepancy between the intermediate clean prediction and the final denoised value of each current-chunk token, DeCoPrune retains high-discrepancy tokens and prunes those with low discrepancy.

Technical Highlights

• Training-Free Method: DeCoPrune is a training-free method, making it easy to integrate into existing autoregressive models.
• Efficient Pruning: In experiments, DeCoPrune achieves over 85% pruning of historical KV tokens while maintaining near-full cache long-range recall.
• Performance Improvement: The generation speed is increased by more than 4 times, significantly outperforming existing compression baselines.

Industry Impact

The release of DeCoPrune marks a significant advancement in the field of autoregressive video generation, particularly in handling long-sequence tasks. Its efficient pruning mechanism and performance improvements provide AI developers with a more powerful tool. Additionally, the method demonstrates the potential of denoising consistency in optimizing model-intrinsic signals, paving the way for future research.

Developer Recommendations

• Integration and Testing: AI developers are advised to integrate DeCoPrune into their existing autoregressive models and conduct extensive testing to evaluate its performance.
• Optimization Strategies: Utilize DeCoPrune's dynamic pruning mechanism to optimize the inference process of the model and improve overall efficiency.
• Explore New Applications: Explore the potential of DeCoPrune in areas such as multimodal generation and real-time interactive systems.

---

Source: Hugging Face Daily Papers (2026-10-05)

— END —

Tags: #Hugging Face Follow #KV Cache Follow #Autoregressive Models Follow #Video Generation Follow #AI Optimization Follow