Research map / Image generation / Editing and style transfer
There and Back Again: On the relation between Noise and Image Inversions in Diffusion Models
Łukasz Staniszewski, Łukasz Kuciński, Kamil Deja
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 |
|---|---|---|---|
| replace first 4% of DDIM inversion steps with forward diffusion (ablation) | DDIM inversion · SDXL latent normality | pixel correlation: 0.16 -> 0.09 | Improved |
| replace first 4% of DDIM inversion steps with forward diffusion (ablation) | DDIM inversion · SDXL latent normality | KL to Gaussian: 0.89 -> 0.38 | Improved |
| replace first 4% of DDIM inversion steps with forward diffusion (ablation) | DDIM inversion · SDXL image reconstruction | PSNR: 27.58 -> 25.68 | Hurt |
| replace first 2% of DDIM inversion steps with forward diffusion (ablation) | DDIM inversion · SDXL image reconstruction | MAE: 0.03 -> 0.03 | No clear effect |
| replace first 4% of DDIM inversion steps with forward diffusion (ablation) | DDIM inversion · SDXL editing | CLIP directional similarity: 75.59 -> 78.91 | Improved |
| replace first 4% of DDIM inversion steps with forward diffusion (ablation) | DDIM inversion · SDXL inversion speed | time per image (s): 7.17 -> 6.86 | Improved |
| use noise-regularized inversion (Pix2Pix-Zero, ReNoise) instead of DDIM | DDIM inversion · SDXL | editability / correlation: corr 0.16 -> 0.15/0.14; time 7.17 -> 22.07/19.86 | No clear effect |
| replace DDIM approximation with reverse DPMSolver | DDIM inversion · SDXL | reconstruction PSNR: 27.58 -> 22.55 (DPMSolver-1) | Hurt |
| apply forward-diffusion fix to StyleAligned inversion | StyleAligned · style transfer from StyleDrop | Style similarity (CSD): 0.690 -> 0.697 | Improved |
| apply forward-diffusion fix to StyleAligned inversion | StyleAligned · style transfer from StyleDrop | Set consistency (DINO): 0.476 -> 0.471 | No clear effect |
| start edits from DDIM latents rather than Gaussian noise (the problem) | DiT/Deepfloyd IF editing · editing diversity | DreamSim diversity: 0.67 -> 0.61 (Deepfloyd IF) | Hurt |
Closest papers by meaning
- An Edit Friendly DDPM Noise Space: Inversion and Manipulations (CVPR 2024 · similarity 0.90)
- A Variational Perspective on Solving Inverse Problems with Diffusion Models (ICLR 2024 · similarity 0.90)
- Divide-and-Conquer Posterior Sampling for Denoising Diffusion priors (NeurIPS 2024 · similarity 0.90)
- Smooth Diffusion: Crafting Smooth Latent Spaces in Diffusion Models (CVPR 2024 · similarity 0.89)
- Gradient-free Decoder Inversion in Latent Diffusion Models (NeurIPS 2024 · similarity 0.89)
- Diffusion Bridge Implicit Models (ICLR 2025 · similarity 0.89)
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.
