RectifID: Personalizing Rectified Flow with Anchored Classifier Guidance

Authors: Zhicheng Sun, Zhenhao Yang, Yang Jin, Haozhe Chi, Kun Xu, Liwei Chen, Hao Jiang, Yang Song, Kun Gai, Yadong Mu

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experimental results on these tasks clearly validate the effectiveness of our approach.
Researcher Affiliation Collaboration 1Peking University, 2Kuaishou Technology, 3University of Electronic Science and Technology of China
Pseudocode Yes Algorithm 1 Anchored Classifier Guidance
Open Source Code Yes Code is available at https://github.com/feifeiobama/Rectif ID. [...] The code to reproduce all results is available at this anonymized link https: //github.com/feifeiobama/Rectif ID, with sufficient instructions in the README file.
Open Datasets Yes For face-centric evaluation, we follow Pang et al. (2024) to evaluate on 20 prompts with the first 200 images from Celeb A-HQ (Liu et al., 2015; Karras et al., 2018) as reference images. For subject-driven generation, we conduct qualitative studies on a subset of examples from the Dream Booth dataset (Ruiz et al., 2023b).
Dataset Splits No The paper describes its method as 'training-free' and evaluates it on specific subsets of datasets (Celeb A-HQ, Dream Booth) rather than defining traditional train/validation/test splits for its own experimental setup.
Hardware Specification Yes The inference time is measured in seconds on an NVIDIA A800. [...] our method takes less than 0.5s per iteration on an NVIDIA A800 GPU and fits on consumer-grade GPUs such as NVIDIA RTX 4080
Software Dependencies No The paper mentions specific models and frameworks like 'Stable Diffusion 1.5' and 'Insight Face library', but does not provide version numbers for general software dependencies like Python, PyTorch, or CUDA.
Experiment Setup Yes We experiment with a frozen piecewise rectified flow (Yan et al., 2024) finetuned from Stable Diffusion 1.5 (Rombach et al., 2022) with 4 equally divided time windows. The number of sampling steps is set to a minimum K = 4 [...] The guidance scale is fixed to s = 1 in quantitative evaluation. [...] The number of iterations is set to N = 100.