Research map / Image generation / Fast diffusion sampling
RectifID: Personalizing Rectified Flow with Anchored Classifier Guidance
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 training-free personalization of rectified flow via anchored classifier guidance, solved as a fixed-point problem using off-the-shelf face/object discriminators (ArcFace, DINOv2) with a convergence guarantee under ideal straightness.
Problem
Personalized image generation that preserves subject identity typically requires finetuning or training on large domain-specific datasets; vanilla classifier guidance needs a noise-aware classifier, limiting use of off-the-shelf discriminators.
Approach
Reformulate classifier guidance for rectified flow as a fixed-point problem on the endpoint z1 (assuming straight trajectories), so clean-image discriminators can be used; anchor the guided trajectory to a reference trajectory to stabilize solving, with a linear convergence guarantee for small guidance scale.
Key results
- RectifID (100 iterations, SD 1.5) reaches identity similarity 0.5930 vs. 0.5806 for InstantID and 0.4778 for IP-Adapter on face-centric evaluation.
- RectifID with 20 iterations takes about 9 s versus 2 s for IP-Adapter, with prompt consistency 0.2995 on SD 1.5.
- Without the anchor, identity similarity drops to 0.1158, compared with 0.5930 for the full anchored method.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add anchored classifier guidance instead of using classifier guidance without an anchor (ablation) | RectifID · CelebA-HQ face-centric personalization | Identity similarity: 0.1158 -> 0.5930 | Helped |
| Add anchored classifier guidance instead of using classifier guidance without an anchor (ablation) | RectifID · CelebA-HQ face-centric personalization | Prompt consistency: 0.2916 -> 0.2933 | Helped |
| Use anchored classifier guidance instead of gradient descent on initial noise (ablation) | RectifID · CelebA-HQ face-centric personalization | Identity similarity: 0.5249 -> 0.5930 | Helped |
| Use anchored classifier guidance instead of gradient descent on initial noise (ablation) | RectifID · CelebA-HQ face-centric personalization | Prompt consistency: 0.2842 -> 0.2933 | Helped |
| Increase the number of iterations from 20 to 100 | RectifID with SD 1.5 · CelebA-HQ face-centric personalization | Identity similarity: 0.4860 -> 0.5930 | Helped |
| Increase the number of iterations from 20 to 100 | RectifID with SD 1.5 · CelebA-HQ face-centric personalization | Prompt consistency: 0.2995 -> 0.2933 | Hurt |
| Replace SD 1.5 with SD 2.1 as the base model | RectifID · CelebA-HQ face-centric personalization | Prompt consistency: 0.2933 -> 0.3151 | Helped |
| Increase the guidance scale from 0 (ablation) | RectifID · CelebA-HQ face-centric personalization | Identity consistency | Helped |
| Increase the number of iterations from 0 (ablation) | RectifID · CelebA-HQ face-centric personalization | Identity consistency | Helped |
| Use a small classifier guidance scale (ablation) | RectifID · CelebA-HQ face-centric personalization | Prompt consistency | Helped |
Surprising findings
- Removing the reference-trajectory anchor collapses identity similarity (0.1158), showing the unanchored fixed-point iteration fails to converge in practice.
- Under 100 iterations, RectifID scores higher identity than training-based InstantID despite requiring no training.
Limitations
- Theoretical guarantee holds only for ideally straight rectified flow and does not extend to general flow models.
- Anchoring works for faces, live subjects and regularly shaped objects, but not for objects with large structural variation such as plush toys.
- Inference time is higher than several training-based baselines.
Experimental setup
Datasets: CelebA-HQ DreamBooth
Benchmarks: CelebA-HQ (200 images, 20 prompts) DreamBooth subset (qualitative)
Metrics: identity similarity (ArcFace) prompt consistency (CLIP) inference time
Baselines: Textual Inversion DreamBooth NeTI Celeb Basis Cross Initialization IP-Adapter PhotoMaker InstantID BLIP-Diffusion Emu2 FastComposer Cones 2
Builds on: classifier guidance rectified flow Universal Guidance DOODL
Closest papers by meaning
- When StyleGAN Meets Stable Diffusion: a W+ Adapter for Personalized Image Generation (CVPR 2024 · similarity 0.85)
- Face2Diffusion for Fast and Editable Face Personalization (CVPR 2024 · similarity 0.85)
- FaceA-Net: Facial Attribute-Driven ID Preserving Image Generation Network (AAAI 2025 · similarity 0.85)
- $\text{ID}^3$: Identity-Preserving-yet-Diversified Diffusion Models for Synthetic Face Recognition (NeurIPS 2024 · similarity 0.85)
- Personalized Representation from Personalized Generation (ICLR 2025 · similarity 0.85)
- Mining your own secrets: Diffusion Classifier Scores for Continual Personalization of Text-to-Image Diffusion Models (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.