SimpleAR: Pushing the Frontier of Autoregressive Visual Generation through Pretraining, SFT, and RL
This work addresses efficient and high-fidelity visual generation for AI applications, though it is incremental as it builds on existing autoregressive methods with optimizations.
The authors tackled high-resolution image generation with an autoregressive model, achieving competitive benchmark scores (0.59 on GenEval and 79.66 on DPG) and reducing generation time to 14 seconds for 1024x1024 images using a 0.5B parameter model.
This work presents SimpleAR, a vanilla autoregressive visual generation framework without complex architecure modifications. Through careful exploration of training and inference optimization, we demonstrate that: 1) with only 0.5B parameters, our model can generate 1024x1024 resolution images with high fidelity, and achieve competitive results on challenging text-to-image benchmarks, e.g., 0.59 on GenEval and 79.66 on DPG; 2) both supervised fine-tuning (SFT) and Group Relative Policy Optimization (GRPO) training could lead to significant improvements on generation aesthectics and prompt alignment; and 3) when optimized with inference acceleraton techniques like vLLM, the time for SimpleAR to generate an 1024x1024 image could be reduced to around 14 seconds. By sharing these findings and open-sourcing the code, we hope to reveal the potential of autoregressive visual generation and encourage more participation in this research field. Code is available at https://github.com/wdrink/SimpleAR.