Research map / Image generation / Autoregressive and latent generation
Co-evolving Representations in Joint Image-Feature Diffusion
Theodoros Kouzelis, Spyros Gidaris, Nikos Komodakis
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 stop-gradient on the clean projected representation target (ablation) | CoReDi-B/2 · ImageNet 256 B/2 200K ablation (DINOv2) | FID: 50.8 -> 24.7 | Improved |
| add batch normalization after the learnable projection (ablation) | CoReDi-B/2 · ImageNet 256 B/2 200K ablation (DINOv2) | FID: 223.9 -> 24.7 | Improved |
| add feature-variance regularization to the learnable projection (ablation) | CoReDi-B/2 · ImageNet 256 B/2 200K ablation (DINOv2) | FID: 37.2 -> 24.7 | Improved |
| unregularized learned projection versus fixed PCA projection (ablation) | ReDi · ImageNet 256 B/2 200K ablation (DINOv2) | FID: 30.9 -> 37.2 | Hurt |
| use learned projection with feature variance regularization instead of fixed PCA | ReDi · ImageNet 256 B/2 200K ablation (DINOv2) | FID: 30.9 -> 24.7 | Improved |
| feature variance versus orthogonality regularization (ablation) | CoReDi-B/2 · ImageNet 256 B/2 200K ablation (DINOv2) | FID: 25.6 -> 24.7 | Improved |
| feature variance versus covariance regularization (ablation) | CoReDi-B/2 · ImageNet 256 B/2 200K ablation (DINOv2) | FID: 25.9 -> 24.7 | Improved |
| raise feature variance regularization weight from 0.5 to 1.5 (ablation) | CoReDi-B/2 · ImageNet 256 B/2 200K ablation (DINOv2) | FID: 25.9 -> 24.3 | Improved |
| set pixel-space representation loss weight to 0.1 instead of larger values (ablation) | CoReDi-L/16 · ImageNet 256 pixel CoReDi-L/16 100K | FID: 33.5 -> 31.5 | Improved |
| use coevolving representation on pixel diffusion | DeCo-L/16 · ImageNet 256 pixel diffusion (no CFG) | FID: 31.3 -> 21.5 | Improved |
| use coevolving representation on pixel diffusion at equal iterations | DeCo-L/16 · ImageNet 256 pixel diffusion (no CFG) | FID: 46.0 -> 31.5 | Improved |
| use CoReDi instead of ReDi with DINOv2 encoder | ReDi · ImageNet 256 B/2 200K, DINOv2 encoder | FID: 30.9 -> 24.7 | Improved |
Closest papers by meaning
- Boosting Generative Image Modeling via Joint Image-Feature Synthesis (NeurIPS 2025 · similarity 0.91)
- Nested Diffusion Models Using Hierarchical Latent Priors (CVPR 2025 · similarity 0.89)
- Würstchen: An Efficient Architecture for Large-Scale Text-to-Image Diffusion Models (ICLR 2024 · similarity 0.89)
- Representation Alignment for Generation: Training Diffusion Transformers Is Easier Than You Think (ICLR 2025 · similarity 0.88)
- End-to-End Training for Unified Tokenization and Latent Denoising (NeurIPS 2026 · similarity 0.88)
- Learning Diffusion Models with Flexible Representation Guidance (NeurIPS 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.
