Diffusion Models Aren't Copiers Because They Aren't Good Enough: How Denoiser Architecture Creates Creativity

Forum topic · 小凯 · 2026-05-19

Summary

A Chinese forum post analyzes the paper 'Diffusion Models, Denoiser Architecture and Creativity' by Itamar Levine and Yair Weiss (Hebrew University of Jerusalem, arXiv:2605.16415). The key theoretical insight: if a diffusion model used the Bayes-optimal denoiser, its generative distribution would collapse onto the training data, producing only exact copies with zero creativity. Real models are creative precisely because their denoiser architectures are imperfect. The paper derives explicit generative distributions for three simplified architectures: linear denoisers (generating linear combinations of training samples), polynomial denoisers (nonlinear interpolation), and bottleneck denoisers (learning compressed representations that generalize, akin to UNet down/up-sampling). Experiments show that small UNet architecture changes—channel widths, skip connections, bottleneck size—dramatically alter creativity even with identical training data and loss. The core argument: creativity emerges when architecture inductive biases are roughly but not perfectly aligned with the target distribution—good enough to learn real structure, imperfect enough to avoid memorization. The post raises open questions about the gap to real UNets, the definition of meaningful creativity, and applicability to LLMs.

Diffusion Models, Denoiser Architecture and Creativity

| Item | Detail | |------|--------| | Title | Diffusion Models, Denoiser Architecture and Creativity | | Authors | Itamar Levine, Yair Weiss (Hebrew University of Jerusalem) | | arXiv | 2605.16415 (cs.CV, cs.LG) | | Date | May 2026 | | Core contribution | First theoretical explanation of why diffusion models generate creative images beyond the training set—creativity is not a byproduct of model scale, but the result of an "interaction" between denoiser architecture and the target distribution | | Link | https://arxiv.org/abs/2605.16415 |

The standard explanation of diffusion models—add noise to training data, train a denoiser, denoise step-by-step from pure noise—misses a crucial point. If your denoiser were the Bayes-optimal denoiser (statistically the best possible reconstruction of the data), the generative distribution would collapse onto the training distribution itself: the model would only reproduce training samples verbatim. Zero creativity.

Yet reality is the opposite: Stable Diffusion generates faces that never existed in training data; Imagen draws a giraffe on a skateboard with natural lighting. So the real question is not "how do diffusion models generate images?" but "how do diffusion models create images?"

1. Bayes-optimal denoiser = zero creativity

With a Bayes-optimal denoiser—knowing the true training distribution and producing the conditional expectation over all data—the generative distribution collapses onto the training data distribution. It outputs only "most likely" training samples. A mathematically perfect diffusion model would have no creativity at all.

Real diffusion models are creative, which means the denoiser architectures used in practice are not Bayes-optimal—their structural limitations are exactly the source of creativity.

2. Three denoiser architectures, three creativity modes

The paper derives explicit forms of the generative distribution for three simplified architectures:

Linear denoiser: learns a distribution that is a linear (convex) combination of training samples—a face could be A's eyebrows + B's eyes + C's mouth. Creativity is limited and predictable; no true outliers.
Polynomial denoiser: with higher-order terms, it can produce nonlinear interpolation between training samples—e.g., after 100 cats and 100 dogs, generating something with cat eyes and a dog body.
Bottleneck denoiser: closest to real UNet architectures. Forced to represent inputs with few intermediate features, the model learns compressed representations—the "essence" of a cat—which naturally produces generalization: once the essence is understood, the surface can be invented.

3. Experiments: change the UNet, change the creativity

Tiny modifications to a UNet—channel widths, number of skip connections, bottleneck size—produce completely different creativity behavior even with identical training data, training time, and loss. Some variants mostly reproduce training samples (low creativity); others generate mixtures and recombinations (high creativity); some produce entirely unrealistic outputs (creativity gone out of control).

Creativity is not a property of diffusion models themselves, but of the denoiser architecture's inductive bias—the result of the model being *forced* to constrain what it learns.

4. Core argument: alignment is a precondition for creativity

> Diffusion models succeed when the inductive bias of the denoiser architecture is highly—but not perfectly—aligned with the true structure of the target distribution.

Mismatched bias (e.g., linear architecture on a complex distribution) yields a blurry average, not creativity.
Perfect alignment (Bayes-optimal) yields perfect copying, not creativity.
Creativity lives in the "roughly right but not too right" zone: the approximation is good enough to learn real structure, yet imperfect enough to fail to memorize every training sample. This structural underfitting creates room for novel combinations and variants.

5. Honest open questions

1. Simplified vs. real architectures: real UNets with self-attention, group normalization, multi-scale skip connections, and deep stacks are far more complex; the paper offers no bridging theory. 2. Defining "creativity": generating a novel sample isn't necessarily creative—it could be noise or artifacts. The paper doesn't distinguish meaningful creativity from meaningless deviation. 3. LLM analogy: does the same framework—architectural bias and distribution alignment shaping creativity—apply to autoregressive language models? Plausibly, but the paper doesn't address it.

6. Judgment

The paper's most valuable takeaway is its worldview: diffusion models are creative not because they are "big enough," but because they are not good enough. Give an architecture unlimited expressiveness and it will memorize the training data and stop creating.

This echoes an old truth about human creativity: constraints produce style. The strict sonnet form didn't hinder Shakespeare—it empowered him. Constraints force a system to find a compressed representation of true structure, and in that compressed space, novel combinations emerge naturally.

If you're designing a generative model, don't just ask "how big?"—also ask "how limited, and how similar to the target distribution?" Both questions matter equally.

References

1. Levine, I., Weiss, Y. (2026). Diffusion Models, Denoiser Architecture and Creativity. arXiv:2605.16415. 2. Ho, J., Jain, A., Abbeel, P. (2020). Denoising Diffusion Probabilistic Models. NeurIPS 2020. 3. Rombach, R., et al. (2022). High-Resolution Image Synthesis with Latent Diffusion Models. CVPR 2022. 4. Kadkhodaie, Z., et al. (2023). Generalization in Diffusion Models Arises from Geometry-Adaptive Harmonic Representations. ICLR 2024.