mlsanityData to 7 Oct 2026

DiffusionPID: Interpreting Diffusion via Partial Information Decomposition

arXiv 2406.05191 ↗First posted 7 Jun 2024cs.CVcited by 16

On the map: Diffusion models → Image generation, text-to-image and editing

Abstract (arXiv)

Text-to-image diffusion models have made significant progress in generating naturalistic images from textual inputs, and demonstrate the capacity to learn and represent complex visual-semantic relationships. While these diffusion models have achieved remarkable success, the underlying mechanisms driving their performance are not yet fully accounted for, with many unanswered questions surrounding what they learn, how they represent visual-semantic relationships, and why they sometimes fail to generalize. Our work presents Diffusion Partial Information Decomposition (DiffusionPID), a novel technique that applies information-theoretic principles to decompose the input text prompt into its elementary components, enabling a detailed examination of how individual tokens and their interactions shape the generated image. We introduce a formal approach to analyze the uniqueness, redundancy, and synergy terms by applying PID to the denoising model at both the image and pixel level. This approach enables us to characterize how individual tokens and their interactions affect the model output. We first present a fine-grained analysis of characteristics utilized by the model to uniquely localize specific concepts, we then apply our approach in bias analysis and show it can recover gender and ethnicity biases. Finally, we use our method to visually characterize word ambiguity and similarity from the model's perspective and illustrate the efficacy of our method for prompt intervention. Our results show that PID is a potent tool for evaluating and diagnosing text-to-image diffusion models.

Key facts. Cited by 16 papers in the map (data through 7 Oct 2026). Placed in Image generation, text-to-image and editing, one of 14 directions in Diffusion models. Builds on 67 of its 85 references that are in the map. First posted on arXiv 7 Jun 2024.

Builds on

67 of 85 references.

Closest ancestors · 19

Same direction, newest first

  1. 2023LEDITS: Real Image Editing with DDPM Inversion and Semantic Guidance2 Jul
  2. 2023Linguistic Binding in Diffusion Models: Enhancing Attribute Correspondence through Attention Map Alignment15 Jun
  3. 2023Grounded Text-to-Image Synthesis with Attention Refocusing8 Jun
  4. 2023The Hidden Language of Diffusion Models1 Jun
  5. 2023Compositional Text-to-Image Synthesis with Attention Map Control of Diffusion Models23 May
  6. 2023Editing Implicit Assumptions in Text-to-Image Diffusion Models14 Mar
  7. 2023Cones: Concept Neurons in Diffusion Models for Customized Generation9 Mar
  8. 2023Adding Conditional Control to Text-to-Image Diffusion Models10 Feb
  9. 2023Attend-and-Excite: Attention-Based Semantic Guidance for Text-to-Image Diffusion Models31 Jan
  10. 2022Uncovering the Disentanglement Capability in Text-to-Image Diffusion Models16 Dec
  11. 2022Training-Free Structured Diffusion Guidance for Compositional Text-to-Image Synthesis9 Dec
  12. 2022Schr"odinger's Bat: Diffusion Models Sometimes Generate Polysemous Words in Superposition23 Nov
7 more
  1. 2022DiffusionDB: A Large-scale Prompt Gallery Dataset for Text-to-Image Generative Models26 Oct
  2. 2022Imagic: Text-Based Real Image Editing with Diffusion Models17 Oct
  3. 2022What the DAAM: Interpreting Stable Diffusion Using Cross Attention10 Oct
  4. 2022Compositional Visual Generation with Composable Diffusion Models3 Jun
  5. 2022Photorealistic Text-to-Image Diffusion Models with Deep Language Understanding23 May
  6. 2022Hierarchical Text-Conditional Image Generation with CLIP Latents13 Apr
  7. 2021GLIDE: Towards Photorealistic Image Generation and Editing with Text-Guided Diffusion Models20 Dec

Other references · 48

  1. 2018BERT: Pre-training of Deep Bidirectional Transformers for Language Understandingcited by 40,204
  2. 2021High-Resolution Image Synthesis with Latent Diffusion Modelscited by 18,694
  3. 2014Microsoft COCO: Common Objects in Contextcited by 17,498
  4. 2016Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localizationcited by 6,690
  5. 2017A Unified Approach to Interpreting Model Predictionscited by 6,488
  6. 2016"Why Should I Trust You?": Explaining the Predictions of Any Classifiercited by 5,040
  7. 2023SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesiscited by 4,290
  8. 2015Learning Deep Features for Discriminative Localizationcited by 2,430
  9. 2022An Image is Worth One Word: Personalizing Text-to-Image Generation using Textual Inversioncited by 2,231
  10. 2017Network Dissection: Quantifying Interpretability of Deep Visual Representationscited by 703
  11. 2022When and why vision-language models behave like bags-of-words, and what to do about it?cited by 603
  12. 2024Evaluating Text-to-Visual Generation with Image-to-Text Generationcited by 465
36 more
  1. 2017Men Also Like Shopping: Reducing Gender Bias Amplification using Corpus-level Constraintscited by 420
  2. 2023TIFA: Accurate and Interpretable Text-to-Image Faithfulness Evaluation with Question Answeringcited by 416
  3. 2023Hard Prompts Made Easy: Gradient-Based Discrete Optimization for Prompt Tuning and Discoverycited by 329
  4. 2022Diffusion Models already have a Semantic Latent Spacecited by 324
  5. 2018Women also Snowboard: Overcoming Bias in Captioning Modelscited by 236
  6. 2023Stable Bias: Analyzing Societal Representations in Diffusion Modelscited by 225
  7. 2021Design Guidelines for Prompt Engineering Text-to-Image Generative Modelscited by 216
  8. 2023JourneyDB: A Benchmark for Generative Image Understandingcited by 203
  9. 2023Davidsonian Scene Graph: Improving Reliability in Fine-grained Evaluation for Text-to-Image Generationcited by 159
  10. 2022CLIP-Dissect: Automatic Description of Neuron Representations in Deep Vision Networkscited by 144
  11. 2022Natural Language Descriptions of Deep Visual Featurescited by 136
  12. 2023Understanding the Latent Space of Diffusion Models through the Lens of Riemannian Geometrycited by 121
  13. 2023Quantifying & Modeling Multimodal Interactions: An Information Decomposition Frameworkcited by 81
  14. 2023Breaking Common Sense: WHOOPS! A Vision-and-Language Benchmark of Synthetic and Compositional Imagescited by 79
  15. 2023Discovering Interpretable Directions in the Semantic Latent Space of Diffusion Modelscited by 60
  16. 2022Testing Relational Understanding in Text-Guided Image Generationcited by 59
  17. 2021Worst of Both Worlds: Biases Compound in Pre-trained Vision-and-Language Modelscited by 52
  18. 2023VisoGender: A dataset for benchmarking gender bias in image-text pronoun resolutioncited by 52
  19. 2023Equivariant Similarity for Vision-Language Foundation Modelscited by 51
  20. 2022DALLE-2 is Seeing Double: Flaws in Word-to-Concept Mapping in Text2Image Modelscited by 50
  21. 2023Intriguing Properties of Data Attribution on Diffusion Modelscited by 47
  22. 2023NoiseCLR: A Contrastive Learning Approach for Unsupervised Discovery of Interpretable Directions in Diffusion Modelscited by 46
  23. 2023Evaluating Data Attribution for Text-to-Image Modelscited by 45
  24. 2022Why is Winoground Hard? Investigating Failures in Visuolinguistic Compositionalitycited by 44
  25. 2023The Journey, Not the Destination: How Data Guides Diffusion Modelscited by 42
  26. 2022Quantifying Societal Bias Amplification in Image Captioningcited by 37
  27. 2023Interpretable Diffusion via Information Decompositioncited by 37
  28. 2019Exposing and Correcting the Gender Bias in Image Captioning Datasets and Modelscited by 36
  29. 2022MultiViz: Towards Visualizing and Understanding Multimodal Modelscited by 36
  30. 2021One Label, One Billion Faces: Usage and Consistency of Racial Categories in Computer Visioncited by 33
  31. 2023Training Data Attribution for Diffusion Modelscited by 24
  32. 2022A Measure of the Complexity of Neural Representations based on Partial Information Decompositioncited by 10
  33. 2023Demystifying Local and Global Fairness Trade-offs in Federated Learning Using Partial Information Decompositioncited by 10
  34. 2022Breaking the Spurious Causality of Conditional Generation via Fairness Intervention with Corrective Samplingcited by 5
  35. 2022Causal effect of racial bias in data and machine learning algorithms on user persuasiveness & discriminatory decision making: An Empirical Studycited by 2
  36. 2024Listen Then See: Video Alignment with Speaker Attentioncited by 2

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2024: 2; 2025: 6; 2026: 8.

16 papers.

Sort:

By area:

  1. 2024textitRevelio: Interpreting and leveraging semantic information in diffusion modelscited by 29
  2. 2024Quantifying Knowledge Distillation Using Partial Information Decompositioncited by 12
  3. 2026How Vision Becomes Language: A Layer-wise Information-Theoretic Analysis of Multimodal Reasoningcited by 7
  4. 2026Towards Understanding Modality Interaction in Multimodal Language Models via Partial Information Decompositioncited by 4
  5. 2026Auditing Language Model Unlearning via Information Decompositioncited by 3
  6. 2025Learning to Match Unpaired Data with Minimum Entropy Couplingcited by 2
  7. 2025On the Separability of Information in Diffusion Modelscited by 2
  8. 2025FAME: Fairness-aware Attention-modulated Video Editingcited by 2
  9. 2025Latent Beam Diffusion Models for Generating Visual Sequencescited by 1
  10. 2025Learning Invariant Graph Representations Through Redundant Informationcited by 1
  11. 2026The Representational Limit of Scalar Interactions: An Interventional Decompositioncited by 1
  12. 2025Decomposing stimulus-specific sensory neural information via diffusion modelscited by 0
4 more
  1. 2026Compensation-free Machine Unlearning in Text-to-Image Diffusion Models by Eliminating the Mutual Informationcited by 0
  2. 2026Self-Captioning Multimodal Interaction Tuning: Amplifying Exploitable Redundancies for Robust Vision Language Modelscited by 0
  3. 2026Concept Guidance: Precise, Training-Free Latent Control for Text-to-Image Generationcited by 0
  4. 2026Interpreting Reasoning of Large Language Models via Partial Information Decompositioncited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2023CONFORM: Contrast is All You Need For High-Fidelity Text-to-Image Diffusion Models
  2. 2023Diffusion Explainer: Visual Explanation for Text-to-image Stable Diffusion
  3. 2023De-Diffusion Makes Text a Strong Cross-Modal Interface
  4. 2024Decoding Diffusion: A Scalable Framework for Unsupervised Analysis of Latent Space Biases and Representations Using Natural Language Prompts
  5. 2023Prompt-Free Diffusion: Taking "Text" out of Text-to-Image Diffusion Models
  6. 2025Emergence and Evolution of Interpretable Concepts in Diffusion Models
  7. 2023Text-to-Image Diffusion Models are Zero-Shot Classifiers
  8. 2023Are Diffusion Models Vision-And-Language Reasoners?