Research map / Architectures and efficiency / RNNs and normalization
ViTok-v2: Scaling Native Resolution Autoencoders to 5 Billion Parameters
Philippe Hansen-Estruch, Jiahui Chen, Vivek Ramanujan, Orr Zohar, Markos Georgopoulos, Animesh Sinha, Ji Hou, Edgar Schönfeld, Felix Juefei-Xu, Sriram Vishwanath, Ali Thabet
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 LPIPS loss to pixel-only training (ablation) | ViTok-v2 5B · ImageNet-1k 256p loss ablation (5B, r=12) | rFID: 5.13 -> 0.72 | Improved |
| add DINOv3 perceptual tile loss (lambda=1000) to pixel-only training (ablation) | ViTok-v2 5B · ImageNet-1k 256p loss ablation (5B, r=12) | rFID: 5.13 -> 0.30 | Improved |
| add DINOv3 loss (lambda=1000) to pixel-only training (ablation) | ViTok-v2 5B · ImageNet-1k 256p loss ablation (5B, r=12) | SSIM: .929 -> .914 | Hurt |
| add SSIM loss to pixel-only training (ablation) | ViTok-v2 5B · ImageNet-1k 256p loss ablation (5B, r=12) | rFID: 5.13 -> 4.87 | Improved |
| use DINOv3 loss instead of LPIPS (ablation) | ViTok-v2 5B · ImageNet-1k 256p loss ablation (5B, r=12) | rFID: 0.72 -> 0.30 | Improved |
| add LPIPS on top of DINOv3 loss (ablation) | ViTok-v2 5B · ImageNet-1k 256p loss ablation (5B, r=12) | rFID: 0.30 -> 0.38 | No clear effect |
| replace 4-block encoder with a single block (ablation) | ViTok-v2 f16 · SSIM | SSIM: .768 -> .753 | No clear effect |
| replace encoder with a linear projection (ablation) | ViTok-v2 f16 · SSIM | SSIM: .768 -> .462 | Hurt |
| scale f32 model width from Bd4-B to Td4-T (ablation) | ViTok-v2 f32 · SSIM | SSIM: .432 -> .872 | Improved |
| scale decoder from B to T at r=48 (ablation) | ViTok-v2 · ImageNet-1k 256p | rFID: B-to-T rFID gap 2.3 at r=12 grows to 12.2 at r=48 | Improved |
| use LayerNorm instead of KL or tanh+noise latent regularization (ablation) | ViTok-v2 AE + 450M flow · ImageNet-1k 256p | gFID: 4.71 -> 4.73 | No clear effect |
| use 5B AE instead of 350M AE for generation (ablation) | 1.2B flow transformer · ImageNet-22k | gFID: 1-2 gFID points better | Improved |
Closest papers by meaning
- Learnings from Scaling Visual Tokenizers for Reconstruction and Generation (ICML 2025 · similarity 0.90)
- When Worse is Better: Navigating the Compression Generation Trade-off In Visual Tokenization (NeurIPS 2025 · similarity 0.87)
- Native-Resolution Image Synthesis (NeurIPS 2025 · similarity 0.86)
- An Image is Worth 32 Tokens for Reconstruction and Generation (NeurIPS 2024 · similarity 0.86)
- VITRIX-UniViTAR: Unified Vision Transformer with Native Resolution (NeurIPS 2025 · similarity 0.86)
- End-to-End Vision Tokenizer Tuning (NeurIPS 2025 · similarity 0.86)
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.
