CoDe: Blockwise Control for Denoising Diffusion Models

Authors: Anuj Singh, Sayak Mukherjee, Ahmad Beirami, Hadi J. Rad

TMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments demonstrate that, despite its simplicity, Co De offers a favorable trade-off between reward alignment, prompt instruction following, and inference cost, achieving a competitive performance against the state-of-the-art baselines. Our code is available at: https://github.com/anujinho/code
Researcher Affiliation Collaboration 1Delft University of Technology, The Netherlands 2Shell Global Solutions International B.V., Amsterdam, The Netherlands 3Massachusetts Institute of Technology, Cambridge MA, USA
Pseudocode Yes Algorithm 1: Co De Algorithm 2: Co De(η)
Open Source Code Yes Our code is available at: https://github.com/anujinho/code
Open Datasets Yes Unless otherwise mentioned, for all experiments, we use a pretrained Stable Diffusion version 1.5 (Rombach et al., 2021) as our base model, which is trained on the LAION-400M dataset (Schuhmann et al., 2021).
Dataset Splits No For quantitative evaluations, we generate 50 images per setting (i.e., prompt-reference image pair) with 500 DDPM steps.
Hardware Specification Yes To achieve this, we have used NVIDIA A100 GPUs with 80GB of RAM.
Software Dependencies No The paper mentions using a 'pretrained Stable Diffusion version 1.5' and the 'CLIP image encoder' but does not specify any version numbers for general software dependencies like Python, PyTorch, or CUDA.
Experiment Setup Yes For quantitative evaluations, we generate 50 images per setting (i.e., prompt-reference image pair) with 500 DDPM steps. ... For the guidance-based methods DPS and UG, the guidance scale is varied between 1 and 50, whereas for the sampling-based methods, Bo N the number of samples N is varied between 2 and 500, while for SVDD and Co De, the number of samples N is varied between 2 and 40.