Compositional Text-to-Image Synthesis with Attention Map Control of Diffusion Models

Authors: Ruichen Wang, Zekang Chen, Chen Chen, Jian Ma, Haonan Lu, Xiaodong Lin

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

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct comprehensive experiments on the publicly available COCO and open-domain datasets, and the results show that our method generates images that are more closely aligned with the given descriptions, thereby improving fidelity and faithfulness.
Researcher Affiliation Collaboration 1OPPO Research Institute 2South China University of Technology 3Rutgers University
Pseudocode Yes Algorithm 1: Denoising Process of Our Method Input: A text prompt p, a trained Box Net B, sets of each parsed entity s token indices {s1, s2, ..., s N}, a trained diffusion model SD Output: Denoised latent z0. 1: for t T, T 1, ..., 1 do 2: boxes B(SD, zt, p, t) 3: for (cx, cy, h, w) in boxes do 4: Convert box to zero-one masks mn 5: Gn Gaussian distribution 2D((cx, cy), h, w) 6: M argmax(Gn) 7: m n (M = n) mn, n = 1, 2..., N unique masks 8: SD SD 9: for each cross attention layer in SD do cross attention mask control 10: Obtain Cross Attention Map C 11: Ci Ci m n i sn, n = 1, 2..., N 12: for each self attention layer in SD do self attention mask control 13: Obtain Self Attention Map S 14: Si Si flatten(m n) i {i|flatten(m n)i = 1}, n = 1, 2..., N 15: zt 1 SD (zt, p, t)
Open Source Code Yes Please refer to https://github.com/OPPOMente-Lab/attention-mask-control.
Open Datasets Yes Specifically, we first train a Box Net applied to the forward process of SD on the COCO dataset (Lin et al. 2014) to predict object boxes for entities with attributes parsed by a constituency parser (Honnibal et al. 2020)." and "We conduct comprehensive experiments on the publicly available COCO and open-domain datasets
Dataset Splits No For evaluation, we construct a new benchmark dataset to evaluate all methods with respect to semantic infidelity issues in T2I synthesis." and "We conduct comprehensive experiments on the publicly available COCO and open-domain datasets". However, it does not explicitly provide training/validation/test splits for their custom benchmark dataset, nor for the COCO dataset beyond mentioning "test split of COCO dataset".
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory) used for running its experiments.
Software Dependencies No The paper mentions software components like spaCy, U-Net, and Stable Diffusion, but it does not specify their version numbers.
Experiment Setup No All the training details and hyper-parameter determination are presented in Appendix A.2. (This indicates that the specific experimental setup details are not in the main text).