Research map / Image generation / Autoregressive and latent generation

Distilled Decoding 1: One-step Sampling of Image Auto-regressive Models with Flow Matching

ICLR 2025 · Image generation › Autoregressive and latent generation · approach: Accelerating autoregressive image generation · 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 Distilled Decoding, which builds deterministic flow-matching trajectories from Gaussian noise to the token outputs of pre-trained image AR models (VAR, LlamaGen) and distills them so generation takes 1 or 2 steps.

Problem

Image autoregressive models generate token by token, which is slow; existing parallel/multi-token decoding methods cannot capture the joint token distribution and fail at one- or two-step generation.

Approach

Use flow matching per token to map Gaussian noise tokens deterministically to the teacher AR model's sampled tokens, forming noise-to-data trajectories along the AR generation process, then train a network to jump from any trajectory point directly to the final sequence, enabling one-step sampling without the original training data.

Key results

• VAR-d16-DD reaches FID 9.94 in 1 step (6.3x speed-up) vs. 4.19 for the 10-step VAR teacher; baselines skip-1/skip-2/onestep* give FID 9.52/40.09/157.5 (skip-1 is better at FID but uses 9 steps).
• LlamaGen-L-DD reaches FID 11.35 in 1 step (217.8x speed-up) vs. 4.11 for the 256-step teacher, while LlamaGen onestep* gives FID 220.2.
• On LAION-COCO text-to-image, LlamaGen-DD with 2 steps gets FID 28.95 vs. 25.70 for the 256-step teacher (92.9x speed-up).

Measured effects reported in the paper

Surprising findings

• Predicting the joint distribution of all tokens in one step (onestep*) yields FID above 150, showing that parallel token prediction cannot capture the AR output distribution.

Limitations

• Performance is bounded by the pre-trained teacher AR model since it is distilled rather than trained from scratch.
• Requires generating a large set of noise-data pairs (1.2M to 3M) with the teacher, which is costly.
• Evaluated only on image AR models; applicability to LLMs with large codebooks and long sequences is left open.

Experimental setup

Datasets: ImageNet-256 LAION-COCO

Benchmarks: ImageNet-256 class-conditional generation LAION-COCO text-to-image (5k prompts)

Metrics: FID IS Precision Recall wall-clock time number of steps

Baselines: VAR skip-1 VAR skip-2 VAR onestep* LlamaGen skip-106 LlamaGen skip-156 LlamaGen onestep* MaskGIT ADM StyleGAN-XL DiT-L/2

Builds on: flow matching rectified flow DPM-Solver consistency models progressive distillation

Closest papers by meaning

• Distilled Decoding 2: One-step Sampling of Image Auto-regressive Models with Conditional Score Distillation (NeurIPS 2025 · similarity 0.95)
• LANTERN: Accelerating Visual Autoregressive Models with Relaxed Speculative Decoding (ICLR 2025 · similarity 0.89)
• Autoregressive Distillation of Diffusion Transformers (CVPR 2025 · similarity 0.89)
• Towards Better & Faster Autoregressive Image Generation: From the Perspective of Entropy (NeurIPS 2025 · similarity 0.89)
• BitDance: Scaling Autoregressive Generative Models with Binary Tokens (NeurIPS 2026 · similarity 0.89)
• Hawk: Leveraging Spatial Context for Faster Autoregressive Text-to-Image Generation (NeurIPS 2025 · similarity 0.88)