Research map / Image generation / Autoregressive and latent generation
Adapting Self-Supervised Representations as a Latent Space for Efficient Generation
Ming Gui, Johannes Schusterbauer, Timy Phan, Felix Krause, Joshua M. Susskind, Miguel Ángel Bautista, Björn Ommer
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 |
|---|---|---|---|
| use pre-trained SSL encoder (DINOv2) instead of no prior (ablation) | RepTok · ImageNet class-cond (10k FID) | gFID: 128.54 -> 20.75 | Improved |
| use no SSL prior instead of SSL (ablation) | RepTok · ImageNet class-cond (10k FID) | PSNR: 19.64 vs 14.94 | Hurt |
| use DINOv2 instead of CLIP as SSL encoder (ablation) | RepTok · ImageNet (10k FID) | gFID: 30.56 -> 20.75 | Improved |
| use DINOv2 instead of MAE as SSL encoder (ablation) | RepTok · ImageNet (10k FID) | gFID: 28.48 -> 20.75 | Improved |
| use DINOv2 instead of CLIP (ablation) | RepTok · ImageNet (10k FID) | rFID: 13.66 -> 7.95 | Improved |
| use single-token latent with MLP-Mixer instead of SiT-XL/2 latent diffusion | RepTok · ImageNet training compute | Total PFlops: 212.5K -> 34.5K | Improved |
| add CFG=1.5 | RepTok · ImageNet | gFID: 3.4 -> 3.22 | Improved |
| scale generator to RepTok-L (516M params) | RepTok · ImageNet | gFID: 3.4 -> 2.06 | Improved |
| fine-tune cls token instead of freezing it (ablation) | RepTok · ImageNet reconstruction | reconstruction quality | Improved |
| increase cosine-similarity weight lambda (ablation) | RepTok · ImageNet | gFID vs PSNR | Mixed |
| use DINOv2 register tokens instead of cls token (frozen) (ablation) | RepTok decoder · ImageNet reconstruction | PSNR / rFID: PSNR 12.59 -> 12.85; rFID 14.13 -> 14.90 | Mixed |
| add register tokens as extra latent tokens (ablation) | RepTok · ImageNet | gFID | Hurt |
Closest papers by meaning
- Generate in Reconstruction Space, Match in Semantic Space: Transport Geometry for One-Step Generation (NeurIPS 2026 · similarity 0.89)
- Can Generative Models Improve Self-Supervised Representation Learning? (AAAI 2025 · similarity 0.89)
- UniTok: a Unified Tokenizer for Visual Generation and Understanding (NeurIPS 2025 · similarity 0.89)
- TokenFlow: Unified Image Tokenizer for Multimodal Understanding and Generation (CVPR 2025 · similarity 0.89)
- When Worse is Better: Navigating the Compression Generation Trade-off In Visual Tokenization (NeurIPS 2025 · similarity 0.89)
- FlexTok: Resampling Images into 1D Token Sequences of Flexible Length (ICML 2025 · similarity 0.88)
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.
