Distilled Decoding (DD): A Novel Approach to Accelerate Autoregressive (AR) Image Generation Models

Autoregressive (AR) models have changed the field of image generation, setting new benchmarks in producing high-quality visuals. These models break down the image creation process into sequential steps, each token generated based on prior tokens, creating outputs with exceptional realism and coherence.

Autoregressive (AR) models have revolutionized the field of image generation, pushing the boundaries of visual realism and coherence. These models operate sequentially, generating each token based on the preceding ones, resulting in outputs of exceptional quality. Researchers have widely employed AR techniques in computer vision, gaming, and digital content creation applications. However, the potential of AR models is often limited by inherent inefficiencies, particularly their slow generation speed, which poses a significant challenge in real-time scenarios.

Among various concerns, a critical aspect that hinders the practical deployment of AR models is their speed. The sequential nature of token-by-token generation inherently limits scalability and introduces high latency during image generation tasks. For instance, generating a 256×256 image using traditional AR models like LlamaGen requires 256 steps, which translates to approximately five seconds on modern GPUs. Such delays hinder their application in scenarios demanding instantaneous results. Moreover, while AR models excel in maintaining the fidelity of their outputs, they face difficulties in meeting the growing demand for both speed and quality in large-scale implementations.

Efforts to accelerate AR models have led to various methods, such as predicting multiple tokens simultaneously or adopting masking strategies during generation. These approaches aim to reduce the required steps but often compromise the quality of the generated images. For example, in multi-token generation techniques, the assumption of conditional independence among tokens introduces artifacts, ultimately undermining the cohesiveness of the output. Similarly, masking-based methods allow for faster generation by training models to predict specific tokens based on others, but their effectiveness diminishes when generation steps are drastically reduced. These limitations highlight the need for a novel approach to enhance AR model efficiency.

A recent research collaboration between Tsinghua University and Microsoft Research has devised a solution to these challenges: Distilled Decoding (DD). This method builds on flow matching, a deterministic mapping that connects Gaussian noise to the output distribution of pre-trained AR models. Unlike conventional methods, DD does not require access to the original training data of the AR models, making it more practical for deployment. The research demonstrated that DD can transform the generation process from hundreds of steps to as few as one or two while preserving the quality of the output. For example, on ImageNet-256, DD achieved a speed-up of 6.3x for VAR models and an impressive 217.8x for LlamaGen, reducing generation steps from 256 to just one.

The technical foundation of DD is based on its ability to create a deterministic trajectory for token generation. Using flow matching, DD maps noisy inputs to tokens to align their distribution with the pre-trained AR model. During training, the mapping is distilled into a lightweight network that can directly predict the final data sequence from a noise input. This process ensures faster generation and provides flexibility in balancing speed and quality by allowing intermediate steps when needed. Unlike existing methods, DD eliminates the trade-off between speed and fidelity, enabling scalable implementations across diverse tasks.

In experiments, DD highlights its superiority over traditional methods. For instance, using VAR-d16 models, DD achieved one-step generation with an FID score increase from 4.19 to 9.96, showcasing minimal quality degradation despite a 6.3x speed-up. For LlamaGen models, the reduction in steps from 256 to one resulted in an FID score of 11.35, compared to 4.11 in the original model, with a remarkable 217.8x speed improvement. DD demonstrated similar efficiency in text-to-image tasks, reducing generation steps from 256 to two while maintaining a comparable FID score of 28.95 against 25.70. The results underline DD’s ability to drastically enhance speed without significant loss in image quality, a feat unmatched by baseline methods.

Several key takeaways from the research on DD include:

In conclusion, with the introduction of Distilled Decoding, researchers have successfully addressed the longstanding speed-quality trade-off that has plagued AR generation processes by leveraging flow matching and deterministic mappings. The method accelerates image synthesis by reducing steps drastically and preserves the outputs’ fidelity and scalability. With its robust performance, adaptability, and practical deployment advantages, Distilled Decoding opens new frontiers in real-time applications of AR models. It sets the stage for further innovation in generative modeling.

Check out the Paper and GitHub Page. All credit for this research goes to the researchers of this project. Also, don’t forget to follow us on Twitter and join our Telegram Channel and LinkedIn Group. Don’t Forget to join our 60k+ ML SubReddit.

Trending: LG AI Research Releases EXAONE 3.5: Three Open-Source Bilingual Frontier AI-level Models Delivering Unmatched Instruction Following and Long Context Understanding for Global Leadership in Generative AI Excellence