Research map / Image generation / Editing and style transfer
Physics-Informed Guided Disentanglement in Generative Networks
Fabio Pizzati, Pietro Cerri, Raoul de Charette
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 differentiable-only parameter estimation instead of full genetic estimation of non-differentiable parameters (ablation) | model-guided disentanglement · clear->rain (drop, nuScenes) | Inception Score: 1.25 -> 1.53 | Improved |
| add physics-model guided disentanglement to MUNIT backbone | MUNIT · clear->rain (drop, nuScenes) | Inception Score: 1.21 -> 1.53 | Improved |
| add physics-model guided disentanglement to MUNIT backbone | MUNIT · gray->color (dirt, WoodScape) | LPIPS diversity: 0.656 -> 0.590 | Hurt |
| add physics-model guided disentanglement to MUNIT backbone | MUNIT · synth->WCS fog | Inception Score: 1.22 -> 1.33 | Improved |
| add physics-model guided disentanglement to MUNIT backbone | MUNIT · clear->snow composite watermark (Synthia) | Inception Score: 1.17 -> 1.19 | Improved |
| finetune PSPNet on model-guided synthetic rain instead of Halder et al. rain rendering | PSPNet · Semantic segmentation on rainy Cityscapes (25 images) | AP: 25.6 -> 27.7 | Improved |
| finetune PSPNet on model-guided synthetic rain instead of original clear-image training | PSPNet · Semantic segmentation on rainy Cityscapes (25 images) | AP: 18.7 -> 27.7 | Improved |
| use neural guidance (DirtyGAN) instead of physical dirt model | gray->color dirt · gray->color (dirt, WoodScape) | SSIM and PSNR | Hurt |
| use neural guidance (DirtyGAN) instead of physical dirt model | gray->color dirt · gray->color (dirt, WoodScape) | quality and diversity | Improved |
| add refraction, shape and thickness variability to simple raindrop model (complex model vs Refract/Gaussian) (ablation) | raindrop guidance model · clear->rain (drop, nuScenes) | GAN metrics | Improved |
| use correct raindrop model instead of wrong model (watermark, dirt, fence) (ablation) | RobotCar raindrop guidance · RobotCar raindrop | FID: 329.17 -> 135.32 | Improved |
| add Disentanglement Guidance with threshold gamma=0.75 (vs 0 or 1) (ablation) | model-guided GAN · clear->rain (drop, nuScenes) | qualitative disentanglement: gamma 0 -> 0.75 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- DRIT — DRIT++: Diverse Image-to-Image Translation via Disentangled Representations (IJCV 2020) · better on 3 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Scaling-Up Disentanglement for Image Translation (ICCV 2021 · similarity 0.88)
- Model-based occlusion disentanglement for image-to-image translation (ECCV 2020 · similarity 0.88)
- Learning Disentangled Representation by Exploiting Pretrained Generative Models: A Contrastive Learning View (ICLR 2022 · similarity 0.88)
- Semi-Supervised StyleGAN for Disentanglement Learning (ICML 2020 · similarity 0.87)
- PhysicsGen: Can Generative Models Learn from Images to Predict Complex Physical Relations? (CVPR 2025 · similarity 0.87)
- Transferring disentangled representations: bridging the gap between synthetic and real images (NeurIPS 2024 · 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.
