Summer may be ending (at least where we are), and we haven’t been to the beach yet, but there’s still time to make a splash!
The scientists here at Anlatan have been busy at work, picking apart our samplers and noise schedules, to see what improvements can be made.
First on the agenda is speed!
All V3 image generation should be a bit faster now. Director Tools and inpainting are the most improved. Vibeless generation, too, enjoys a bit of extra speed on top of our across-the-board improvements.
We’ve also found faster defaults; 23 steps of our new Karras schedule performs well with Euler Ancestral sampling. The results are competitive with our old 28-step schedule in blind testing. Of course the limits for free Opus generations didn’t change.
New Karras schedule? Yes, that brings us to the schedule fixes.
We’ve updated Karras, Exponential and Polyexponential noise schedules to spend their steps better.
Compared to the native schedule:
- As step count increases: steps are spent where they’re most needed.
- Karras is a body specialist.
- Exponential is similar to Karras, but with a step-spacing that helps multistep samplers.
- Polyexponential is a specialist at small features like fingers.
We hope you’ll enjoy trying the new schedules, and exploring how low you can push the step count.
Next let’s see what’s new with multistep samplers.
DPM++ 2M no longer exhibits rainbow artifacting:
DPM++ 2M, 18-step native
The problem was especially pronounced on the old Karras schedule:
DPM++ 2M, 12-step Karras
We also welcome in a new sampler, DPM++ 2M SDE.
Like its cousin DPM++ 2M: it excels at producing sharp images at low step counts.
Its key difference is that it has extra error-correction capability, like Euler Ancestral and DPM++ SDE.
DPM++ 2M SDE, 10-step Karras
We introduce Variety Boost, a new feature to improve the diversity of samples.
Variety Boost improves diversity of samples
Variety Boost is closely-related to Prompt Guidance. You’re probably already familiar with Prompt Guidance: setting it higher improves relevance and coherence, and setting it too high deepfries your image. But Guidance has another impact, often overlooked: it reduces the variation in images. You may have found that some prompts will place your character on a white background every time, use a limited set of poses, or draw clothing the same way each time, even when your prompt left room for interpretation.
Variety Boost seeks to get the benefits of Guidance without the downsides, by enabling it later in the generation process[0], after the broad composition and body shape have been decided. This still allows the face and clothing details to be relevant, and still activates in time to ensure coherence of limbs and hands.
Note: Variety Boost means your negative prompt will only be used after the body shape has been decided.
Have a play with these new settings, and show us what you can make.
Let’s all finish summer without regrets!
References
[0] Variety Boost implements the idea presented in Kynkäänniemi et al’s paper, Applying Guidance in a Limited Interval Improves Sample and Distribution Quality in Diffusion Models: https://arxiv.org/abs/2404.07724
