ComicCraft-LCM

This is an experimental version of ComicCraft, incorporating Latent Consistency Models (LCM).

The intention was to do this in a way that it would still work with most of the regular samplers and still benefit from being able to use less Sampling steps (for example 6, instead of the typical setting of more than 20) to obtain high quality results much faster.

The results are not identical to those of the original ComicCraft model, but this might still be interesting to test with, specially for users of applications that may have limitations with using LoRAs.

Regarding the trigger words for the model, it should be the same as the original ComicCraft model, but some styles are weaker.

As mentioned, most of the standard samplers will work, but A1111 doesn’t yet have a LCMScheduler (as far as I understand), so the full benefits from LCM can’t be used yet. The DPM++ 2M and 3M samplers might have some issues at lower sampling steps (blurry or “overburned” images), but they can still be useable from 8 steps onwards (for example 10 or 12, depending on the case).

From my tests, using a configuration like this should be ok:

• Sampling steps: at least 6

• Samplers: Euler, Euler a and UniPC will work well without doing too much. Others may need higher sampling steps

• CFG scale: between 1 and 2

I tested some styles and character LoRAs and they still work ok, but some may need some adjustments. Try lowering the LoRA weight slightly.

ADetailer and other extensions work well and given the speedups of reducing the number of sampling steps, it will still be useful for faster image generation.

The images in the showcase have a variety of samplers and sampling steps used, so they can be a reference. Hires. fix was not used for any of them, ADetailer was used in a few of them.

I cannot test everything on my own, so if you find out anything interesting, let me know.