Loss-Guided Diffusion Models for Plug-and-Play Controllable Generation
Authors: Jiaming Song, Qinsheng Zhang, Hongxu Yin, Morteza Mardani, Ming-Yu Liu, Jan Kautz, Yongxin Chen, Arash Vahdat
ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate LGD-MC on various synthetic and real-world experiments, such as a one-dimensional mixture of Gaussian (Sec. 6.1), image super-resolution (Sec. 6.2), controllable motion synthesis (Sec. 6.3). |
| Researcher Affiliation | Collaboration | 1NVIDIA Corporation 2Georgia Institute of Technology |
| Pseudocode | No | The paper describes its algorithms verbally and with computational graphs (Figure 5) but does not include structured pseudocode or an explicitly labeled algorithm block. |
| Open Source Code | No | The paper does not include an explicit statement about releasing its source code or a direct link to a code repository for the methodology described. |
| Open Datasets | Yes | In this experiment, we use the pre-trained unconditioned 256 256 Image Net diffusion model from Open AI guided diffusion models (Dhariwal & Nichol, 2021). |
| Dataset Splits | Yes | We follow the approach in Anonymous (2022), where we use the DDIM sampler (Song et al., 2021a) with η = 1.0 (i.e., the DDPM sampler) for 100 steps over the 50k Image Net validation set. |
| Hardware Specification | Yes | To support our claims in Sec. 4.3 on compute and memory, we evaluate the memory consumption and average time to perform one iteration over image super-resolution on the same NVIDIA RTX 3090 GPU. |
| Software Dependencies | No | The paper mentions using standard deep learning frameworks and models (e.g., Image Net diffusion models), but it does not specify exact version numbers for programming languages, libraries, or other software dependencies. |
| Experiment Setup | Yes | We follow the approach in Anonymous (2022), where we use the DDIM sampler (Song et al., 2021a) with η = 1.0 (i.e., the DDPM sampler) for 100 steps over the 50k Image Net validation set. For all experiments, we use the same motion diffusion model (Tevet et al., 2022) and the 100-step DDIM sampler, with rt = σt/ p 1 + σ2 t . |