Manifold Preserving Guided Diffusion
Authors: Yutong He, Naoki Murata, Chieh-Hsin Lai, Yuhta Takida, Toshimitsu Uesaka, Dongjun Kim, Wei-Hsiang Liao, Yuki Mitsufuji, J Zico Kolter, Ruslan Salakhutdinov, Stefano Ermon
ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experiments show that MPGD is efficient and effective for solving a variety of conditional generation applications in low-compute settings, and can consistently offer up to 3.8 speed-ups with the same number of diffusion steps while maintaining high sample quality compared to the baselines. Our code is available via the project page here. |
| Researcher Affiliation | Collaboration | 1Carnegie Mellon University , 2Sony AI , 3Sony Group Corporation , 4Stanford University |
| Pseudocode | Yes | Algorithm 1 MPGD for pixel diffusion models, Algorithm 2 MPGD for latent diffusion models, Algorithm 3 g M: On-Manifold Guidance |
| Open Source Code | Yes | Our code is available via the project page here. |
| Open Datasets | Yes | We evaluate our approach using the FFHQ 256x256 (Karras et al., 2019) and Image Net 256x256 (Deng et al., 2009) datasets. ...We test all methods with the pretrained diffusion model for the Celeb A-HQ 256x256 dataset provided by Yu et al. (2023)... For reference style images and text prompts, we randomly created 1000 conditioning pairs, using images from Wiki Art (Saleh & Elgammal, 2015) and prompts from Parti Prompts (Yu et al., 2022) dataset. |
| Dataset Splits | No | The paper specifies evaluation on existing 'test sets' but does not explicitly describe the full training, validation, and test dataset splits that were performed for their own experimental setup, as their method is training-free and uses pre-trained models. |
| Hardware Specification | Yes | All experiments are conducted on a single NVIDIA Ge Force RTX 2080 Ti GPU. ...All the samples are generated on a single NVIDIA RTX3090 GPU. ...All the samples are generated on a single NVIDIA A100 GPU. |
| Software Dependencies | No | The paper does not provide specific software dependencies with version numbers (e.g., Python, PyTorch, CUDA versions) necessary for replication. |
| Experiment Setup | Yes | The parameter η is set to 0.5. The weight parameter scheduling is based on the implementation of DPS. The guidance weight hyperparameters for all of MPGD w/o proj., MPGD-AE, and MPGD-Z are 20,10,5 for DDIM steps 20, 50, 100 respectively. The weights for DPS is 0.3 as their default, for MCG is 100.0 and for LGD is 0.05 for the best empirical results we obtain. We follow the super-resolution setting in the LGD paper for its additional weight scheduling. The number of Monte Carlo samples for LGD is set to 10. |