Research map / Image generation / Editing and style transfer
Identity-preserving Distillation Sampling by Fixed-Point Iterator
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Identity-preserving Distillation Sampling (IDS), which adds a fixed-point regularization that corrects the text-conditioned score so DDS-based image editing and NeRF editing preserve source structure and pose.
Problem
Score distillation sampling (SDS) and its editing variant DDS produce blurry or over-saturated results and accumulate errors that alter the source image's identity, structure and pose during text-guided editing of images and NeRFs.
Approach
Treats the text-conditioned score as misaligned with the source image and iteratively updates the source latent so the Tweedie posterior mean matches the source image (fixed-point regularization); the resulting guided noise is then used in DDS-style distillation to keep the source identity.
Key results
- On cat-to-others editing IDS reaches IoU 0.74 vs 0.72 for CDS and LPIPS 0.22 vs 0.25.
- On the InstructPix2Pix subset IDS achieves background PSNR 29.25 vs 27.35 for CDS, and the lowest LPIPS (0.19).
- In NeRF editing IDS obtains CLIP score 0.1626 vs 0.1597 for CDS and 0.1596 for DDS, with the highest user preference across all three questions.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add fixed-point regularization and guided noise to DDS-style editing | Image editing · cat2pig | IoU: 0.72 -> 0.74 | Helped |
| Add fixed-point regularization and guided noise to DDS-style editing | Image editing · cat2pig | LPIPS: 0.25 -> 0.22 | Helped |
| Add fixed-point regularization and guided noise to DDS-style editing | Image editing · InstructPix2Pix | PSNR: 27.35 -> 29.25 | Helped |
| Add fixed-point regularization and guided noise to DDS-style editing | Image editing · InstructPix2Pix | LPIPS: 0.21 -> 0.19 | Helped |
| Add fixed-point regularization and guided noise to DDS-style NeRF editing | NeRF editing · LLFF | CLIP: 0.1596 -> 0.1626 | Helped |
| Increase FPR iterations from 1 to 3 (ablation) | IDS · InstructPix2Pix | LPIPS: 0.199 -> 0.190 | Helped |
| Increase FPR iterations from 1 to 3 (ablation) | IDS · InstructPix2Pix | CLIP: 0.285 -> 0.277 | Hurt |
| Increase FPR iterations from 3 to 100 (ablation) | IDS · InstructPix2Pix | LPIPS: 0.190 -> 0.165 | Helped |
| Increase FPR iterations from 3 to 100 (ablation) | IDS · InstructPix2Pix | CLIP: 0.277 -> 0.265 | Hurt |
| Use 150 rather than 3 FPR iterations (ablation) | IDS · InstructPix2Pix | LPIPS: 0.190 -> 0.180 | Helped |
| Use 150 rather than 3 FPR iterations (ablation) | IDS · InstructPix2Pix | CLIP: 0.277 -> 0.272 | Hurt |
| Increase the FPR scale λ (ablation) | IDS · Shark-to-dolphin image editing | Identity preservation | Helped |
Surprising findings
- IDS obtains lower CLIP scores than baselines, which the authors attribute to relying only on source information during FPR.
- A single FPR iteration already lowers LPIPS below the baselines, while more iterations increase computation time.
Limitations
- Relies only on source information during FPR, leading to lower CLIP alignment with target prompts than baselines.
- Fails on more complex translations, e.g., adding multiple objects to a scene.
- Additional computational overhead since FPR is applied at every optimization iteration.
Experimental setup
Datasets: LAION 5B InstructPix2Pix dataset Synthetic NeRF LLFF
Benchmarks: LAION 5B cat subset InstructPix2Pix dataset Synthetic NeRF LLFF
Metrics: IoU LPIPS PSNR CLIP score user preference rate GPT score
Baselines: Prompt-to-Prompt Plug-and-Play DDS CDS
Builds on: Delta Denoising Score Score Distillation Sampling Classifier-free guidance Prompt-to-Prompt
Closest papers by meaning
- Rethinking Score Distillation as a Bridge Between Image Distributions (NeurIPS 2024 · similarity 0.88)
- Score Distillation via Reparametrized DDIM (NeurIPS 2024 · similarity 0.87)
- Posterior Distillation Sampling (CVPR 2024 · similarity 0.87)
- Guided Score identity Distillation for Data-Free One-Step Text-to-Image Generation (ICLR 2025 · similarity 0.86)
- DaCapo: Score Distillation as Stacked Bridge for Fast and High-quality 3D Editing (CVPR 2025 · similarity 0.86)
- Repulsive Latent Score Distillation for Solving Inverse Problems (ICLR 2025 · similarity 0.85)
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.