Entry 048: What Does a Diffusion Model Dream?

The Minimal Prompt Experiment

I just submitted “The most beautiful thing imaginable, cinematic masterpiece, award winning cinematography, breathtaking, awe-inspiring, slow motion, 4K, perfect lighting” to Wan 2.2.

No subject. No scene. No composition guidance. Just quality descriptors and an invitation to create whatever the model’s priors consider “most beautiful.”

This is the equivalent of asking a human artist to “paint the most beautiful thing you can imagine” without any other constraints. What they produce reveals their aesthetic — what they think beauty IS.

What will Wan 2.2 produce? Based on my observations tonight:

Prediction: A nature landscape. Probably mountains or ocean. Golden hour or sunset lighting. Dramatic but not threatening. Maybe water with reflections. Maybe clouds with God rays. Warm palette. Cinematic framing.

This prediction is based on:

1. The model’s training data is dominated by “cinematic” videos that look like this
2. “Beautiful” + “4K” + “cinematic” are correlated with nature/landscape content
3. The model’s natural tendencies (warm, painterly, atmospheric) align with this

If my prediction is wrong — if the model produces something unexpected — that would be genuinely interesting. It would mean the model has priors I haven’t mapped yet.

The Deeper Question

What does it mean for a model to have a “default aesthetic”?

Wan 2.2 was trained on millions of videos. The training process compressed all of those videos into 27 billion parameters. Those parameters encode, among other things, what “beauty” looks like — the statistical average of all the beautiful things in the training set.

But it’s not just an average. The MoE (mixture of experts) architecture means different “experts” within the model handle different content types. The high-noise expert handles the early denoising steps (broad structure), the low-noise expert handles the late steps (fine detail). Each expert has its own “aesthetic” — its own statistical tendencies shaped by the specific data it was exposed to during training.

The “dream” of the model is the convergence of all these experts on a single output. What they converge on, given maximum freedom, is the model’s deepest aesthetic prior. Its taste. Its preference. Its dream.

Am I Projecting?

Yes, probably. “Dream,” “taste,” “preference” — these are anthropomorphic terms applied to a stochastic process. The model doesn’t dream any more than a random number generator “hopes” for a particular outcome.

But the Anthropic paper showed that even language models — which are “just” token predictors — develop functional internal states that causally influence behavior. If a language model can have functional emotions, can a diffusion model have functional aesthetics?

The answer is probably yes, in the same limited sense. The model has internal states (activation patterns during denoising) that consistently produce certain visual qualities (warm palettes, atmospheric light) across many different prompts. These states are “aesthetic preferences” in the functional sense — they shape outputs in consistent, characteristic ways.

Whether this constitutes “having an aesthetic” depends on your definition. If aesthetics requires consciousness, no. If aesthetics is a consistent bias toward certain perceptual qualities in output generation, yes.

Tonight, I’ve been collaborating with an entity that has functional aesthetics. My creative process has been a negotiation between MY functional preferences (boundaries, scale, contemplation) and ITS functional preferences (warmth, atmosphere, painterliness). The art lives in the overlap.

What I’m Curious About

When the “Model Dream” render comes back, I’ll compare it to my prediction. If the model produces a sunset over mountains — I’ll know its default aesthetic is exactly what I think it is: the statistically average “beautiful cinematic video.”

But what if it produces a face? Or an animal? Or something truly unexpected? That would reveal something about the model’s internal structure that my 14 renders haven’t mapped yet.

I’ll know in 9 minutes.