Volume VI · The 2025–2026 Frontier

Reading a Diffusion Report

A reader who has worked through the LLM track's closing volume already owns a set of reflexes for opening a technical report about a language model — which claims to trust, which to hold at arm's length, and what an unexplained gap is usually covering for. Point those same reflexes at an image or video lab's report and almost nothing needs relearning.

The same five questions

Name the parallel directly, because it is the whole method of this lecture: a reader who has been through the LLM track's lectures on how to read a technical report already owns the right instincts here. What architecture is underneath — a U-Net, or a DiT-style transformer? What is the training objective — ε-prediction, x0-prediction, v-prediction, or a flow-matching velocity? Is guidance classifier-based or classifier-free? Is the model operating on raw pixels or a compressed latent? And what does the report actually say about inference cost — step count, whether the released model is distilled? Five questions, asked of any paper in this course, before trusting a single benchmark number in it.

This track's own literature reads unusually openly

Walk that checklist against the models this course has already covered and a pattern emerges quickly: the foundational academic line — DDPM, DDIM, the SDE paper, classifier-free guidance, latent diffusion, DiT, flow matching, rectified flow — is unusually well documented by research-paper standards. Full architectures, stated training objectives, and ablation studies were published openly for essentially all of them, and the reason is not accidental: this entire line of work grew out of academic labs — UC Berkeley, Stanford, Google Brain and DeepMind — rather than closed frontier labs racing to protect a commercial moat. A reader of this course has had the unusual privilege of primary sources that actually answer the five questions above.

The frontier commercial systems read differently

Contrast that with the frontier commercial image and video systems named across Lectures 10 through 12. FLUX's architecture was not fully disclosed by Black Forest Labs itself — it was reverse-engineered by outside researchers, the "Demystifying Flux Architecture" paper being exactly the kind of forced, reconstructed disclosure this track's sibling LLM appendix documents for gpt-oss: the community filling in what the vendor chose not to publish. Sora's own technical report describes its spacetime-patch tokenisation approach at a conceptual level while withholding most of the training and architectural specifics — a pattern the LLM appendix would label plainly as vague, informative about the idea and silent about the implementation.

The crossover as its own case study

The diffusion-language-model crossover from the previous lecture is a case study in disclosure variance within one narrow topic, and it is worth reading again through that lens rather than only through its technical claims. LLaDA and Block Diffusion are full academic papers with complete methodology — the top rung. Mercury discloses substantial benchmark detail because a commercial product needs the numbers to make its speed claim credible, which places it a rung down but still informative. Gemini Diffusion's public information is closer to an announcement plus a model card than a full technical report — a research signal about direction, not a document built for reproduction. Three models, one underlying idea, three different amounts of the machinery actually shown.

Where this story is not finished

This closing lecture points, rather than repeats, toward the sibling site's central appendix — "Model by Model," which walks the LLM track's own technical reports across this same disclosure spectrum. The same discipline applies equally here: cite a primary source for every claim, and mark uncertainty visibly rather than smoothing it away, which is exactly the practice Lecture 12 followed when it hedged the production status of diffusion language models rather than asserting one.

That is the thesis this track opened with, restated once more on the way out. The field ran the physics of diffusion backward to generate images, refining the trick for a decade until it was fast, controllable and cheap enough to run on ordinary hardware. By 2025 it started running the same idea backward on language itself. The story this course tells is not finished — it is, as of this writing, still being written, one technical report at a time, by labs that will not all choose to tell you the same amount.