Research map / Image generation / Autoregressive and latent generation
AToken: A Unified Tokenizer for Vision
Jiasen Lu, Liangchen Song, Mingze Xu, Byeongjoo Ahn, Yanjun Wang, Chen Chen, Afshin Dehghan, Yinfei Yang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add video training (Stage 2) to image-only tokenizer (ablation) | AToken-So/C (800M) · ImageNet reconstruction | rFID: 0.26 -> 0.25 | Improved |
| add 3D training (Stage 3) after image and video (ablation) | AToken-So/C (800M) · ImageNet reconstruction | rFID: 0.25 -> 0.21 | Improved |
| add 3D training (Stage 3) (ablation) | AToken-So/C (800M) · TokenBench video reconstruction | rFVD: 3.63 -> 3.01 | Improved |
| use smaller Base (192M) model instead of So400m (800M) under multimodal expansion (ablation) | AToken · ImageNet rFID across stages | rFID: 0.323 -> 0.483 | Hurt |
| use Gram matrix loss instead of GAN loss for reconstruction (ablation) | AToken transformer tokenizer · ImageNet reconstruction | rFID | Improved |
| add video stage before 3D stage for generation tokenizer (ablation) | AToken-So/C with Lightning-DiT · ImageNet generation (continuous) | gFID: 1.62 -> 1.88 | Hurt |
| add 3D stage (Stage 3) after video stage (ablation) | AToken-So/C with Lightning-DiT · ImageNet generation (continuous) | gFID: 1.88 -> 1.56 | Improved |
| quantize latents with FSQ (discrete variant) | AToken-So · ImageNet reconstruction | rFID: 0.21 -> 0.38 | Hurt |
| replace Oryx-ViT with frozen AToken-So/C encoder in SlowFast-LLaVA-1.5-7B | SlowFast-LLaVA-1.5-7B · image understanding (7 benchmarks) | accuracy: RW-QA 67.5 -> 68.8; OCRBench 76.4 -> 74.5 | Mixed |
| replace Oryx-ViT with frozen AToken-So/C encoder in SlowFast-LLaVA-1.5-7B | SlowFast-LLaVA-1.5-7B · video understanding | accuracy: VideoMME 63.9 -> 64.5; LongVideoBench 62.5 -> 60.6 | Mixed |
Closest papers by meaning
- TokenFlow: Unified Image Tokenizer for Multimodal Understanding and Generation (CVPR 2025 · similarity 0.91)
- UniFlow: A Unified Pixel Flow Tokenizer for Visual Understanding and Generation (ICLR 2026 · similarity 0.90)
- UniTok: a Unified Tokenizer for Visual Generation and Understanding (NeurIPS 2025 · similarity 0.89)
- OmniTokenizer: A Joint Image-Video Tokenizer for Visual Generation (NeurIPS 2024 · similarity 0.89)
- Rosetta Stone For Unified MLLMs: A Unified Tokenizer to Decipher Understanding and Generation (CVPR 2026 · similarity 0.89)
- End-to-End Vision Tokenizer Tuning (NeurIPS 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
