Experiment

Diffusion Schedulers: A Quick Comparison

This note compares DDIM, Euler, and DPM++ schedulers on image quality, speed, and prompt consistency for the same checkpoint and seed set.

Goal

Identify a scheduler default that balances quality and generation speed for daily experimentation.

Test Setup

  • Model: SDXL base checkpoint
  • Prompt set: 20 prompts across portrait, product, and environment categories
  • Seed policy: fixed seed list for cross-scheduler comparison
  • Step counts tested: 20, 30, 40

Early Readout

Euler produced faster iterations for ideation, while DPM++ gave more consistent fine detail at higher step counts.

Next Step

Add a prompt adherence scoring rubric and compare results across two CFG ranges.

  • diffusion
  • sampling
  • image-generation