Research map / Video / Video generation
Control-DINO: Feature Space Conditioning for Controllable Video Diffusion
Edoardo Dominici, Thomas Deixelberger, Konstantinos Vardis, Markus Steinberger
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 |
|---|---|---|---|
| train with appearance augmentation (real + styled data) instead of real data only (ablation) | Control-DINO · Tanks and Temples transfer (stride 8) | CLIP similarity: 0.4408 -> 0.4717 | Improved |
| train with appearance augmentation instead of real data only (ablation) | Control-DINO · Tanks and Temples transfer (stride 8) | Aesthetic quality: 0.5533 -> 0.5610 | Mixed |
| keep spatial mixing of conditioning with transformer hidden state (ablation) | Control-DINO · DL3DV140 architecture ablation | FID: 177.80 -> 71.67 | Improved |
| use raw 384-d DINO features instead of style-invariant PCA projection (ablation) | Control-DINO · DL3DV140 architecture ablation | FID: 89.14 -> 71.67 | Improved |
| use full features instead of PCA tail drop (k=8) (ablation) | Control-DINO · DL3DV140 architecture ablation | FID: 156.67 -> 71.67 | Improved |
| condition on DINO features instead of Canny or depth | CogVideoX-5B-I2V ControlNet · DL3DV140 (in domain, real) | FID: 37.65 -> 20.23 | Improved |
| condition on DINO features instead of depth/Canny at longer stride | Control-DINO · Tanks and Temples transfer (stride 16) | COLMAP registration rate: 0.5918 -> 0.9932 | Improved |
| condition on DINO features instead of Canny at short stride | Control-DINO · Tanks and Temples transfer (stride 8) | COLMAP registration rate: 0.9770 -> 0.9530 | Hurt |
| use 2D DINO features instead of Canny edges | Control-DINO · ScanNet++ video-from-3D | PSNR: 19.82 -> 27.93 | Improved |
| derive conditioning from untextured 3D (voxel) instead of 2D video features | Control-DINO · ScanNet++ video-from-3D | PSNR: 27.93 -> 22.80 | Hurt |
| remove color from Concerto point-cloud features (inpainted) (ablation) | Control-DINO Concerto · ScanNet++ video-from-3D | PSNR: 21.19 -> 19.66 | Hurt |
| inject ControlNet into more transformer layers (4 vs more) (ablation) | Control-DINO · ScanNet++ reconstruction | FID: 39.82 -> 36.67 | Improved |
Closest papers by meaning
- VD3D: Taming Large Video Diffusion Transformers for 3D Camera Control (ICLR 2025 · similarity 0.89)
- Diffusion Renderer: Neural Inverse and Forward Rendering with Video Diffusion Models (CVPR 2025 · similarity 0.87)
- Ctrl-Adapter: An Efficient and Versatile Framework for Adapting Diverse Controls to Any Diffusion Model (ICLR 2025 · similarity 0.87)
- AC3D: Analyzing and Improving 3D Camera Control in Video Diffusion Transformers (CVPR 2025 · similarity 0.87)
- Universal Few-shot Spatial Control for Diffusion Models (NeurIPS 2025 · similarity 0.87)
- DORSal: Diffusion for Object-centric Representations of Scenes $\textit{et al.}$ (ICLR 2024 · similarity 0.87)
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.
