liveBook · Manning

DPM++ 2M

As explained in Local Image Generation

DPM++ 2M is a second-order multistep diffusion sampling algorithm for updating the denoising trajectory.

DPM++ 2M is a second-order multistep diffusion sampler: think of it as a careful barber that removes noise in measured passes rather than hacking away blindly. In Stable Diffusion, its results depend not only on the sampler but also on the schedule type and step count. section 6.1.2

Use DPM++ 2M as a practical comparison tool:

  • Keep the prompt, checkpoint, seed, dimensions, CFG, and schedule type fixed when comparing samplers. section 6.2.2
  • Start with a modest step count, then increase it if the image looks unstable or sampler differences seem pronounced. section 6.2.2
  • Treat schedule type as a creative control: it changes how much noise is removed at each step. section 6.1.2
  • Do not assume DPM++ 2M is always best; visual preference can change with the prompt and settings. section 6.1.2

A quick sampler test

Generate “a sand castle on the beach” at 512×512 with Deliberate_v5, 10 steps, Karras scheduling, seed 1337, and CFG 7. Compare DPM++ 2M with DDIM and Euler. At 10 steps, DPM++ 2M may look especially sharp; at 30 steps, the three results become much more alike. section 6.2.2

Figure 6.9 Increasing the CFG scale shows limitations in higher order samplers.

Compare the sampler outputs after raising CFG to 10; the figure illustrates a limitation for DPM++ 2M at only 10 steps. section 6.2.2

How Local Image Generation covers it

How deep it goes
Explains sampler behavior through visual comparisons and settings, but stops short of deriving the second-order algorithm internally.
Examples
Compares sand-castle generations at 10 and 30 steps, plus a high-CFG test using fixed prompts and seeds.
What it assumes
Assumes basic Stable Diffusion generation settings while explaining sampler and schedule choices through concrete images.
What you can do after
Evaluate sampler settings by generating controlled image batches and adjusting steps, schedules, or CFG for preferred results.
How it connects
Connects DPM++ 2M with diffusion samplers, schedule types, sampling steps, CFG scale, DDIM, Euler, and Stable Diffusion.

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Local Image Generation. The section links point to the parts of the book it draws on.