Research map / Image generation / Autoregressive and latent generation
On the Diffusibility of High-Dimensional Latents
Chao Feng, Zhiyang Xu, Bowei Chen, Yuanjun Xiong, Xiyao Wang, Jui-Hsien Wang, Richard Zhang, Zhe Lin, Andrew Owens, Yijun Li
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 |
|---|---|---|---|
| finetune encoder with reconstruction loss (keep v-prediction) (ablation) | DINOv2-L RAE · text-to-image | GenEval: 37.63 -> 30.96 | Hurt |
| finetune encoder with reconstruction loss (keep v-prediction) (ablation) | DINOv2-L RAE · COCO-30k | FID: 18.34 -> 29.97 | Hurt |
| switch v-prediction to x0-prediction (ablation) | FT-DINOv2-L · text-to-image | GenEval: 30.96 -> 39.48 | Improved |
| switch v-prediction to x0-prediction (ablation) | FT-DINOv2-L · COCO-30k | FID: 29.97 -> 16.80 | Improved |
| switch v-prediction to x0-prediction (ablation) | MAE-B-RAE · text-to-image | GenEval: 36.17 -> 40.89 | Improved |
| switch v-prediction to x0-prediction (ablation) | MAE-B-RAE · COCO-30k | FID: 22.20 -> 17.24 | Improved |
| switch v-prediction to x0-prediction (ablation) | MAE-B-RAE · DPG-Bench | score: 67.54 -> 73.90 | Improved |
| finetune DINOv2-L for reconstruction plus x0-prediction vs frozen DINOv2-L v-prediction | DINOv2-L · COCO-30k | FID: 18.34 -> 16.80 | Improved |
| compress to 32-d latent instead of x0 in 1024-d | FT-DINOv2-L · text-to-image | GenEval / FID: GenEval 40.99 vs 39.48; FID 17.64 vs 16.80 | Mixed |
| reconstruction finetuning reduces effective dimensionality | DINOv2-L · ImageNet val | R90: 672 -> 129 | Hurt |
| reconstruction finetuning | DINOv2-L · ImageNet-1k linear probe | accuracy (%): 84.5 -> 63.9 | Hurt |
Closest papers by meaning
- Improved Baselines with Representation Autoencoders (NeurIPS 2026 · similarity 0.89)
- Improving the Diffusability of Autoencoders (ICML 2025 · similarity 0.88)
- Reconstruction vs. Generation: Taming Optimization Dilemma in Latent Diffusion Models (CVPR 2025 · similarity 0.88)
- Improving Progressive Generation with Decomposable Flow Matching (NeurIPS 2025 · similarity 0.87)
- Cross Flow: One-Step Generation Across Latent and Pixel Spaces (NeurIPS 2026 · similarity 0.86)
- Unleashing Diffusion Transformers for Visual Correspondence by Modulating Massive Activations (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.
