Research map / Image generation / Autoregressive and latent generation
Collaborative Decoding Makes Visual Auto-Regressive Modeling Efficient
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 Collaborative Decoding (CoDe), which splits VAR next-scale generation between a large drafter (VAR-d30, 2B) for early small scales and a small refiner (VAR-d16/0.3B) for late large scales, with specialized fine-tuning, to speed up inference and cut KV-cache memory on ImageNet 256.
Problem
Visual Auto-Regressive (VAR) next-scale prediction produces very long token sequences at the final scales, causing high KV-cache memory use and computational redundancy, which slows inference.
Approach
Use a large VAR model to draft the low-frequency coarse scales and a small VAR model to refine the high-frequency fine scales, since large scales need fewer parameters and the generation patterns of small and large scales are exclusive; each model is then fine-tuned only on its assigned scales (CSE loss for drafter, KD for refiner).
Key results
- VAR-CoDe (N=8) achieves 1.7x speedup and about 50% memory reduction with FID 1.98 vs 1.95 for VAR-d30.
- VAR-CoDe (N=6) reaches 2.9x speedup (about 50 images/s at batch 64 on L20) with FID 2.27.
- VAR-CoDe (N=6) is about 15x faster than DiT-XL/2 at comparable quality and 60x faster than LlamaGen-XXL.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace later VAR-d30 decoding steps with a VAR-d16 refiner (N=8) | VAR-d30 · ImageNet1K-256 | FID: 1.95 -> 1.98 | Hurt |
| Replace later VAR-d30 decoding steps with a VAR-d16 refiner (N=8) | VAR-d30 · ImageNet1K-256 | speedup: 1.0x -> 1.7x | Helped |
| Use 6 rather than 8 drafting steps in collaborative decoding (ablation) | VAR-CoDe · ImageNet1K-256 | speedup: 1.7x -> 2.9x | Helped |
| Use 6 rather than 8 drafting steps in collaborative decoding (ablation) | VAR-CoDe · ImageNet1K-256 | FID: 1.98 -> 2.27 | Hurt |
| Add specialized fine-tuning to the drafter and refiner (ablation) | VAR-CoDe · ImageNet1K-256 | FID: 2.42 -> 2.27 | Helped |
| Add specialized fine-tuning to the drafter and refiner (ablation) | VAR-CoDe · ImageNet1K-256 | FID: 2.26 -> 2.11 | Helped |
| Add specialized fine-tuning to the drafter and refiner (ablation) | VAR-CoDe · ImageNet1K-256 | FID: 2.10 -> 1.98 | Helped |
| Add specialized fine-tuning to the drafter and refiner (ablation) | VAR-CoDe · ImageNet1K-256 | FID: 1.99 -> 1.94 | Helped |
| Fine-tune only the largest three scales of VAR-d16 (ablation) | VAR-d16 · ImageNet-256 | FID: 3.30 -> 21.93 | Hurt |
Surprising findings
- Fine-tuning the large-scale tokens alone collapsed small-scale modeling (FID rose from 3.30 to 21.93), showing strong interference between scales.
- A 2B drafter matched a 0.3B model at the final scale, indicating large scales need few parameters.
Limitations
- Speedup is smaller at low batch sizes (1.6x at batch 1) due to low GPU utilization.
- Requires fine-tuning both drafter and refiner; training-free version is weaker (FID 2.42 at N=6).
Experimental setup
Datasets: ImageNet-1K
Benchmarks: ImageNet-1K 256x256
Metrics: FID IS Precision Recall latency throughput memory consumption speedup
Baselines: VAR-d30 VAR-d24 VAR-d20 LlamaGen-XXL AiM-XL MAR-B DiT-XL/2
Builds on: VAR knowledge distillation LlamaGen MAR
Closest papers by meaning
- LANTERN: Accelerating Visual Autoregressive Models with Relaxed Speculative Decoding (ICLR 2025 · similarity 0.91)
- Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction (NeurIPS 2024 · similarity 0.89)
- Accelerating Auto-regressive Text-to-Image Generation with Training-free Speculative Jacobi Decoding (ICLR 2025 · similarity 0.89)
- ZipAR: Parallel Autoregressive Image Generation through Spatial Locality (ICML 2025 · similarity 0.88)
- Memory-Efficient Visual Autoregressive Modeling with Scale-Aware KV Cache Compression (NeurIPS 2025 · similarity 0.88)
- Hawk: Leveraging Spatial Context for Faster Autoregressive Text-to-Image Generation (NeurIPS 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.