Offline as of August 2023. Fun toy, not something to maintain indefinitely. (And the serverless GPU landscape keeps changing.)

See: diffusion-grid.silverthorn.blog

See also: @bsilverthorn/diffusion-grid-banana

It's a simple visualization toy for diffusion model trajectories.

It visualizes "paths not taken" during the denoising process.

It's a toy because it was built for playing around and building a bit of intuition. It's not a tool for research or prompt development.

It looks like this:

When a diffusion model produces an image, that output is influenced by two sources of randomness: its random initial position in latent space, and any random decisions¹ made throughout the denoising process.

We can imagine different trajectories, producing different images, all starting from the same random initial position. If we focus on one of these trajectories, we can also imagine different possible trajectories branching from each point on that trunk, i.e., sets of trajectories that are all identical up to a given timestep. Trajectories branching from later timesteps² will be more similar than trajectories branching from earlier timesteps.

Diffusion Grid shows images produced by several trajectories branching from each of several different timesteps, e.g.:

  graph LR;
      S[Initial] --> T0i[t = 1000]
      T0i --> E0i["t = …"]
      E0i --> T0j[t = 500]
      T0j --> E0j["t = …"]
      E0j --> I0[Trunk Image]
      T0j --> E1j["t = …"]
      E1j --> I1[Branch Image]
      T0j --> E2j["t = …"]
      E2j --> I2["⋮"]
      T0i --> E3i["t = …"]
      E3i --> T3j[t = 500]
      T3j --> E3j["t = …"]
      E3j --> I3[Branch Image]
      T0i --> E4i["t = …"]
      E4i --> T4j[t = 500]
      T4j --> E4j["t = …"]
      E4j --> I4["⋮"]

I was curious about the emergence of image structure within the reverse trajectory. What kind of "choices" are made earlier? Later?

One way to visualize that emergence is to look at multiple trajectories with differing similarity, as described above, which also turns out to be visually interesting and kind of fun to play around with.

(If you have pointers to any of these topics in the literature, please pass them along.)