* Merge Lab series *

Documentation:

This is a "SD XL" merge model and finished with a PCM Lora in my little computer (GTX 1060-6Gb/16Gb Ram) using comfyui and an extension "mecha_merger" (a model merger with focus on low memory footprint) and pruned on automatic 1111

With this model you can create consistent images at 8 steps with normal cfg

How to use it.

Sampling method

1. DPM++SDE (slow, but precise, best choice for low steps)

2. DDIM (good for general compositions, it can use for low steps)

3. DPM++2m (good for general compositions, it can use for low steps)

4. Heun++2 (for comfyui)

5. CFG++ (good for finalization using img2img)

6. LMS (karras) (good for all composition and finalization on img2img, it can use for low steps)

8. DEIS (good for general compositions, it can use for low steps)

9. UniPC

10. IPNDM (good when using img2img)

11. EULER-A ( Karras OR Align Your Steps, CFG 4 OR 5, optional-> for best results "FreeU")

Schedule type

karras (good results)

SGM Uniform (for fast inference, use it with 8-10 steps)

Align Your Steps (AYS) for 10 steps

Sampling steps

8 - 10 for a low step composition with good results when used in the 4 - 5 CFG scale

Styles

sensitive to "fluffyrock" tags, non furry artists, semi-realistic and "photographic" compositions

1. no style -> In general, semi-realistic or realistic ("photographic")

2. prompt: realistic -> semi-realistic

3. negative prompt: low quality, cartoon, toony, 3d -> semi-realistic to realistic

4. prompt: realistic and negative prompt: low quality, cartoon, toony, 3d -> realistic ("photographic")