Stylus: Automatic Adapter Selection for Diffusion Models

Authors: Michael Luo, Justin Wong, Brandon Trabucco, Yanping Huang, Joseph E. Gonzalez, zhifeng Chen, Ruslan Salakhutdinov, Ion Stoica

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

Reproducibility Variable Result LLM Response
Research Type Experimental To evaluate Stylus, we developed Stylus Docs, a curated dataset featuring 75K adapters with pre-computed adapter embeddings. In our evaluation on popular Stable Diffusion checkpoints, Stylus achieves greater CLIP/FID Pareto efficiency and is twice as preferred, with humans and multimodal models as evaluators, over the base model.
Researcher Affiliation Collaboration Michael Luo1 Justin Wong1 Brandon Trabucco2 Yanping Huang3 Joseph E. Gonzalez1 Zhifeng Chen3 Ruslan Salakhutdinov2 Ion Stoica1 1UC Berkeley 2CMU MLD 3Google Deepmind
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks (clearly labeled algorithm sections or code-like formatted procedures).
Open Source Code Yes Stylus is already open-source, with all experiments made available to individuals online on Github.
Open Datasets Yes We evaluate Stylus over a cross product of two datasets, Microsoft COCO [22] and Parti Prompts [53]
Dataset Splits Yes We evaluate COCO 2014 validation dataset, with 10K sampled prompts
Hardware Specification Yes We launched 16 replicas of Stylus and Stable Diffusion on 8 A100-80GB GPUs for 4 weeks to generate images for evaluation.
Software Dependencies Yes We assess Stylus against Stable-Diffusion-v1.5 [40] as the baseline model. ... In our experiments, these improved descriptions were generated by Gemini Ultra [43] ... In our experiments, we create embeddings from Open AI s text-embedding-3-large model [21, 30]. ... In our implementation, we choose Gemini 1.5, with a 128K context window, as the composer s LLM
Experiment Setup Yes Our image generation process integrates directly with Stable-Diffusion Web UI [1] and defaults to 35 denoising steps using the default DPM Solver++ scheduler [26]. To replicate high-quality images from existing users, we enable high-resolution upscaling to generate 1024x1024 from 512x512 images, with the default latent upscaler [17] and denoising strength set to 0.7.