Concept Algebra for (Score-Based) Text-Controlled Generative Models

Authors: Zihao Wang, Lin Gui, Jeffrey Negrea, Victor Veitch

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate the idea with examples using Stable Diffusion. ... We now work through some examples testing if this subspace structure does indeed exist, and whether it enables manipulation of concepts. ... We compare this with directly prompting the model, and with concept composition [e.g., Du+21; Liu+21]. ... Raters consistently favored images produced by concept algebra, as highlighted in table 1.
Researcher Affiliation Collaboration Zihao Wang1, Lin Gui1, Jeffrey Negrea2, and Victor Veitch1,2,3 1Department of Statistics, University of Chicago 2Data Science Institute, University of Chicago 3Google Research
Pseudocode Yes Algorithm 1 Concept Manipulation through Projection ... Algorithm 2 Find Subspace Basis ... Algorithm 3 Find Subspace Mask
Open Source Code No The paper does not provide a direct link or explicit statement about the open-source release of their own code for the described methodology.
Open Datasets No The paper refers to 'training data' in a general sense (e.g., 'training data for large-scale generative models') and mentions fine-tuning a diffusion model, but it does not specify any publicly available datasets used for their experiments with concrete access information (e.g., links, DOIs, or formal citations).
Dataset Splits No The paper does not provide specific details on training, validation, and test dataset splits, such as percentages or sample counts.
Hardware Specification No The paper does not provide specific details about the hardware used to run the experiments (e.g., GPU models, CPU types, or memory specifications).
Software Dependencies No The paper mentions using Stable Diffusion and Dreambooth, and references external software in [Liu+22], but does not provide specific version numbers for any software dependencies.
Experiment Setup Yes First, we fine-tune the diffusion model using Dreambooth, applying a learning rate of 5e 6 and setting the number of steps to 800. ... During sampling, the original prompt is utilized in the first 20% of timesteps to better retain the content. ... To ensure image fidelity, we exclusively employ the score function sdreambooth( a sks toy ) for guiding the denoising process for the last 6% of the denoising steps.