ImageFolder🚀: Autoregressive Image Generation with Folded Tokens

     

Updates

• (2024.11.14) Code will be released in two weeks (company approval in progress).
• (2024.10.03) We are working on advanced training of ImageFolder tokenizer.
• (2024.10.01) Repo created. Code and checkpoints will be released soon.

Ablation (updating)

ID	Method	Length	rFID ↓	gFID ↓	ACC ↑
🔶1	Multi-scale residual quantization (Tian et al., 2024)	680	1.92	7.52	-
🔶2	+ Quantizer dropout	680	1.71	6.03	-
🔶3	+ Smaller patch size K = 11	265	3.24	6.56	-
🔶4	+ Product quantization & Parallel decoding	265	2.06	5.96	-
🔶5	+ Semantic regularization on all branches	265	1.97	5.21	-
🔶6	+ Semantic regularization on one branch	265	1.57	3.53	40.5
🔷7	+ Stronger discriminator	265	1.04	2.94	50.2
🔷8	+ Equilibrium enhancement	265	0.80	2.60	58.0

🔶1-6 are already in the released paper, and after that 🔷7+ are advanced training settings used similar to VAR (gFID 3.30).

Generation

Visualization of Decomposed Token

Acknowledge

We would like to thank the following repositories: LlamaGen, VAR and ControlVAR.

Citation

If our work assists your research, feel free to give us a star ⭐ or cite us using

@misc{li2024imagefolderautoregressiveimagegeneration,
      title={ImageFolder: Autoregressive Image Generation with Folded Tokens}, 
      author={Xiang Li and Hao Chen and Kai Qiu and Jason Kuen and Jiuxiang Gu and Bhiksha Raj and Zhe Lin},
      year={2024},
      eprint={2410.01756},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2410.01756}, 
}