AI Video Made Crystal Clear · chapter 2: How Video Models Work
Forward and reverse diffusion on one frame
2026-09-20
The forward chain is a fixed recipe used only in training; the reverse chain is what runs when you press Generate. After Weng (2021, updated 2024).
Below: the paragraphs from the book that build this idea, then the diagram itself (Figure 2.2), and a recap. About a minute of reading.
Second, train a network to run the recipe backwards. Show it a frame at noise level t and ask it to guess what noise was added. The practical trick, again from Weng, is to "predict noise from the input at time step t" rather than the clean picture, which turns out to be an easier target. Do this across millions of clips and the network becomes a specialist in one question: given this static, which speck does not belong?
Generation is then the reverse recipe run on fresh noise. Start from pure static, ask the network for its guess, remove a little, ask again. Figure 2.2 shows both directions on one frame.
Recap
- The idea: The forward chain is a fixed recipe used only in training; the reverse chain is what runs when you press Generate.
- The picture: Figure 2.2, from chapter 2 ("How Video Models Work") of AI Video Made Crystal Clear.
- Go deeper: the chapter builds this step by step, with recipes and sources at the end.
This diagram is one of many in AI Video Made Crystal Clear.
Every chapter opens with the gist, draws the hard ideas, and ends with recipes and sources.
Get the book
