mlsanityData to 14 Oct 2026

Catastrophic Compositional Generation: Why Vanilla Diffusion Models Fail to Extrapolate

arXiv 2606.23920 ↗First posted 22 Jun 2026cs.LG cs.AIcited by 1

On the map: Diffusion models → Diffusion foundations and sampling

Abstract (arXiv)

The task of compositional generation involves using a conditional generative model, trained only on a subset of the possible conditions, to produce samples from compositionally-defined target distributions such as a geometric combination of the source distributions. In this work, we argue that this task is often infeasible for vanilla conditional diffusion models: we conjecture that no inference-time technique can efficiently produce samples from the target distribution in certain well-motivated settings. This idea is supported by theory-guided generalization arguments and carefully-designed experiments on both synthetic and realistic data. In particular, while recent methods such as Feynman-Kac correction reduce inference-time approximation error, our results show that score estimation error has a more catastrophic effect on performance when the target distribution is out-of-distribution with respect to the sources, highlighting the need for a different approach to this task.

Key facts. Cited by 1 paper in the map (data through 14 Oct 2026). Placed in Diffusion foundations and sampling, one of 14 directions in Diffusion models. Builds on 33 of its 51 references that are in the map. First posted on arXiv 22 Jun 2026.

Builds on

33 of 51 references.

Closest ancestors · 11

Same direction, newest first

  1. 2025DriftLite: Lightweight Drift Control for Inference-Time Scaling of Diffusion Models25 Sep
  2. 2025Feynman-Kac Correctors in Diffusion: Annealing, Guidance, and Product of Experts4 Mar
  3. 2025Mechanisms of Projective Composition of Diffusion Models6 Feb
  4. 2023SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis4 Jul
  5. 2023Common Diffusion Noise Schedules and Sample Steps are Flawed15 May
  6. 2022Classifier-Free Diffusion Guidance26 Jul
  7. 2021High-Resolution Image Synthesis with Latent Diffusion Models20 Dec
  8. 2021Diffusion Models Beat GANs on Image Synthesis11 May
  9. 2021Improved Denoising Diffusion Probabilistic Models18 Feb
  10. 2020Score-Based Generative Modeling through Stochastic Differential Equations26 Nov
  11. 2020Denoising Diffusion Probabilistic Models19 Jun

Other references · 22

  1. 2015U-Net: Convolutional Networks for Biomedical Image Segmentationcited by 19,713
  2. 2023DINOv2: Learning Robust Visual Features without Supervisioncited by 7,974
  3. 2022Flow Matching for Generative Modelingcited by 5,702
  4. 2017FiLM: Visual Reasoning with a General Conditioning Layercited by 2,333
  5. 2023Stochastic Interpolants: A Unifying Framework for Flows and Diffusionscited by 939
  6. 2022Compositional Visual Generation with Composable Diffusion Modelscited by 624
  7. 2022Interventional Causal Representation Learningcited by 142
  8. 2023Learning Linear Causal Representations from Interventions under General Nonlinear Mixingcited by 92
  9. 2023Object-Centric Slot Diffusioncited by 90
  10. 2023SlotDiffusion: Object-Centric Generative Modeling with Diffusion Modelscited by 85
  11. 2018Marginal Singularity, and the Benefits of Labels in Covariate-Shiftcited by 65
  12. 2023Concept Algebra for (Score-Based) Text-Controlled Generative Modelscited by 59
10 more
  1. 2022Causal Structure Learning: a Combinatorial Perspectivecited by 46
  2. 2023General Identifiability and Achievability for Causal Representation Learningcited by 34
  3. 2024Score-based Causal Representation Learning: Linear and General Transformationscited by 27
  4. 2023From Identifiable Causal Representations to Controllable Counterfactual Generation: A Survey on Causal Generative Modelingcited by 22
  5. 2021Out-of-Distribution Generalization in Kernel Regressioncited by 21
  6. 2020Learning Robust Models Using The Principle of Independent Causal Mechanismscited by 18
  7. 2024Linear Causal Representation Learning from Unknown Multi-node Interventionscited by 14
  8. 2026Enhanced Diffusion Sampling: Efficient Rare Event Sampling and Free Energy Calculation with Diffusion Modelscited by 13
  9. 2024Automated Discovery of Pairwise Interactions from Unstructured Datacited by 8
  10. 2025Virtual Cells: Predict, Explain, Discovercited by 7

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2026: 1.

1 paper.

  1. 2026Error-Corrected Inference-Time Scaling for Imperfect Diffusion Modelscited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2023Compositional Abilities Emerge Multiplicatively: Exploring Diffusion Models on a Synthetic Task
  2. 2025CoInD: Enabling Logical Compositions in Diffusion Models
  3. 2025Improving Compositional Generation with Diffusion Models Using Lift Scores
  4. 2026Test-Time Compositional Generalization in Diffusion Models via Concept Discovery
  5. 2023Reduce, Reuse, Recycle: Compositional Generation with Energy-Based Diffusion Models and MCMC
  6. 2022Sampling is as easy as learning the score: theory for diffusion models with minimal data assumptions
  7. 2026TILT: Improving Compositional Generation in Diffusion Models with a Model-Intrinsic Reward
  8. 2024Conditional sampling within generative diffusion models