CADS: Unleashing the Diversity of Diffusion Models through Condition-Annealed Sampling
Authors: Seyedmorteza Sadat, Jakob Buhmann, Derek Bradley, Otmar Hilliges, Romann M. Weber
ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we rigorously evaluate the performance of CADS on various conditional diffusion models and demonstrate that CADS boosts output diversity without compromising quality. |
| Researcher Affiliation | Collaboration | 1ETH Z urich, 2Disney Research|Studios |
| Pseudocode | Yes | The precise algorithms for CADS and Dynamic CFG are detailed in Algorithms 1 and 2, along with the pseudocode for the annealing schedule and noise addition in Figure 15. |
| Open Source Code | No | The paper states 'Our work builds upon publicly available datasets and the official implementations of the pretrained models cited in the main text.' and refers to the 'official codebase of LDM', but does not explicitly state that the code for CADS itself is open-sourced or provided. |
| Open Datasets | Yes | We consider four conditional generation tasks: class-conditional generation on Image Net (Russakovsky et al., 2015) with Di T-XL/2 (Peebles & Xie, 2022), pose-to-image generation on Deep Fashion (Liu et al., 2016) and SHHQ (Fu et al., 2022), identity-conditioned face generation with ID3PM (Kansy et al., 2023), and text-to-image generation with Stable Diffusion (Rombach et al., 2022). |
| Dataset Splits | Yes | For evaluating the Stable Diffusion model, we sample 1000 random captions from the COCO validation set (Lin et al., 2014) and generate 10 samples per prompt. |
| Hardware Specification | No | The paper mentions using the 'ETH Z urich Euler computing cluster' but does not provide specific hardware details such as GPU models, CPU types, or memory specifications. |
| Software Dependencies | No | The paper mentions using 'official implementations of the pretrained models' and refers to the 'official codebase of LDM', but does not list specific versions for software dependencies such as programming languages, libraries, or frameworks (e.g., Python, PyTorch, TensorFlow, CUDA versions). |
| Experiment Setup | Yes | Table 13 includes the sampling hyperparameters used to create each table in the main text. |