Research map / Image generation / Autoregressive and latent generation

HeatKV: Head-tuned KV-cache Compression for Visual Autoregressive Modeling

NeurIPS 2026 · Image generation › Autoregressive and latent generation · approach: Next-scale and token autoregressive images · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes HeatKV, a head-and-scale-specific static KV-cache pruning schedule for Visual Autoregressive models, ranked by calibration-set attention (S-CAS), with a greedy early-pruning scheme that enforces the memory budget at every layer.

Problem

VAR image generation models have very large KV-caches (up to ~10 GB per 1024x1024 image) that cap batch size; existing VAR KV-cache compressors use coarse binary per-layer or per-head budgets.

Approach

Rank each (layer, head, source-scale) triple by the attention mass that future scales place on it (S-CAS) from a small offline calibration set, then prune head-scale pairs in that order to meet a memory budget, enforcing the budget after every layer rather than only at scale boundaries.

Key results

• On Infinity-2B at 10% budget, HeatKV reaches FID 1.67, LPIPS 0.073, PSNR 25.40 on MS-COCO, beating the best baseline (ScaleKV) at 20% budget (FID 1.94, PSNR 24.93).
• At 4% budget on Infinity-2B, HeatKV achieves HPSv2.1 30.45, above all baselines at 20% budget.
• On Infinity-8B at 20% budget, HeatKV gives FID 1.04 and PSNR 27.44 versus ScaleKV's 1.44 and 25.60.

Surprising findings

• Calibration with a single prompt performs about as well as with 10 prompts, indicating head rankings are robust to calibration set size.
• Pruning naively at scale boundaries causes transient cache over-allocation in VARs, which the paper's per-layer early-pruning avoids.

Limitations

• Relies on stability of attention patterns across prompts; may break if they are not stable.
• Assumes all tokens within a scale are equally important, so low-importance tokens may be retained.
• Gains on GenEval are small and often within noise of baselines.

Experimental setup

Datasets: MS-COCO 2017 validation (evaluation) 10 calibration prompts generated by Qwen3-VL-2B-Instruct

Benchmarks: MS-COCO 2017 GenEval HPSv2.1

Metrics: FID LPIPS PSNR GenEval overall score HPSv2.1

Baselines: StreamingLLM SnapKV HACK ScaleKV Full Cache

Builds on: Infinity ScaleKV HACK StreamingLLM SnapKV

Closest papers by meaning

• Memory-Efficient Visual Autoregressive Modeling with Scale-Aware KV Cache Compression (NeurIPS 2025 · similarity 0.92)
• VL-Cache: Sparsity and Modality-Aware KV Cache Compression for Vision-Language Model Inference Acceleration (ICLR 2025 · similarity 0.87)
• VideoMLA: Low-Rank Latent KV Cache for Minute-Scale Autoregressive Video Diffusion (NeurIPS 2026 · similarity 0.87)
• SmallKV: Small Model Assisted Compensation of KV Cache Compression for Efficient LLM Inference (NeurIPS 2025 · similarity 0.87)
• Mixing Importance with Diversity: Joint Optimization for KV Cache Compression in Large Vision-Language Models (ICLR 2026 · similarity 0.87)
• KV-COBRA: KV Cache Compression via Co-Optimized Bit-Rank Allocation (NeurIPS 2026 · similarity 0.86)