无需视觉 tokenizer，北大PRA解锁自回归图像生成潜力，135M模型性能反超1.9B基线

In deep learning, model scale is often seen as a guarantee of performance. However, the PRA method proposed by Peking University breaks this conventional wisdom. According to a report by JiQiZhiXin, the PRA model with only 135M parameters outperforms a 1.9B-parameter baseline model in autoregressive image generation tasks. The parameter count differs by more than 14 times, yet PRA achieves the overtaking—a shocking result that has prompted researchers to rethink which components in model architecture are truly necessary.

Traditional autoregressive image generation typically relies on visual tokenizers (e.g., VQ-VAE) to discretize continuous pixels into discrete token sequences. PRA's key innovation is to completely eliminate this preprocessing step, directly performing autoregressive modeling on the raw pixel space. This not only simplifies the pipeline but also greatly reduces the parameter count, as the tokenizer itself often introduces a large number of parameters. PRA's success suggests that, after removing the tokenizer, a more compact model can learn more generalizable representations.

PRA's core hypothesis is that visual tokenizers may be a computational bottleneck, limiting the model's ability to capture precise spatial information. By operating directly on pixels, PRA avoids information loss from discretization while forcing the model to use its limited parameters more efficiently. The report states that PRA beats the 1.9B baseline on multiple standard image generation benchmarks; specific metrics are not disclosed, but the term 'overtaking' implies significant improvement.

This evidence preliminarily supports the feasibility of the 'de-tokenizer' approach. However, note that the report comes from JiQiZhiXin, a secondary summary, and is a release of results from PKU itself. The full experimental details, ablation studies, and comparison baselines (e.g., LLM-based image generation) of the original paper have not been publicly verified. Therefore, while the results are exciting, the confidence level as an independent fact should be reserved.

There are currently two main technical routes for autoregressive image generation: one, represented by DALL-E and Parti, relies on powerful visual tokenizers and large language models; the other attempts to bypass tokenizers and model pixels directly. PRA clearly belongs to the latter, but its performance overtaking indicates that this route may have been underestimated.

If PRA's conclusions are widely replicated, it would mean that researchers can move away from reliance on tokenizers and extremely large models, instead focusing on more efficient architecture design. This is particularly attractive for resource-constrained academic labs and startups. At the same time, it could push new-generation image generation models toward lighter weight and easier deployment.

However, caution is needed: a single breakthrough does not equal a paradigm shift. The 1.9B baseline may not be the most representative, and whether PRA can maintain its advantage at higher resolutions and on more complex datasets remains unknown. The final outcome of the route debate depends on independent validation and horizontal comparison by more teams.

First, the report does not specify the exact metrics (e.g., FID, IS) used for PRA's surpassing of the baseline, nor does it detail the baseline model configuration. The lack of standardized comparison leaves the precise meaning of 'overtaking' to be clarified. Second, the computational cost of autoregressive pixel modeling may explode with increasing image resolution; it is unclear whether PRA has addressed this.

Furthermore, as a method proposed by the PKU team, its reproducibility needs to be verified through open-source code and detailed hyperparameter disclosure. Finally, does extreme parameter efficiency come at the cost of scalability? Can PRA maintain its advantage when model size continues to grow? The answers to these questions will determine whether PRA is a flash in the pan or a foundation for the future.