VERSION 3.0

Version 3 is out!

This one focuses on blindingly fast generation, photographic realism and dramatic impact. It incorporates a bit of Big Lust and a mix of self-trained LORAs, but mostly I have been tinkering with the clip and text encoders, aiming for decent production at very low step numbers.

The model is extremely resource-efficient, and can generate very nice images with a sweet spot in precisely 3 LCM steps, although such low numbers can occasionally produce some anatomical horrors. Switching to 20 LCM steps and above improves quality in complex descriptions and is still very fast.

As the previous version, it combines pretty well with LORAs and embeddings from different flavors of Stable Diffusion XL, Pony and Illustrious.

The first 5 examples use no lora or embedding, and the first 4 are obtained with just 3 steps (the barbarian image is obtained with 20 steps to avoid strangeness with the sword grip), just to show the capabilities of the model.

For resolutions above 1280x1280, you should use some hi-res fix, for example with a first generation at 1024x1024 refined at 1536x1536.

As before, beware of the NSFW capabilities of the model, and prompt carefully.

Please experiment, have fun, give thumbs up if you like it, and don't hesitate to give feedback!

OPTIMAL SETTINGS: This version is optimized for LCM sampling only.

I recommend:

CLIP Skip: 1, for maximum realism.

Ultrafast: LCM sampler / CFG 0.9-3 / 3 steps / noise shift 1 - 1.2 

Improved quality: LCM sampler / CFG0.9-3 / 20+ steps / noise shift 1-1.2

VERSION 2.0

This photorealistic model is a mixed merge of dozens of checkpoints and LORAs from the past couple of years, including quite a few of the latter directly trained by myself. Unfortunately, as I did not plan ahead to release it and I have been adapting it for so long, I have not kept track of everything included in it, so I apologize in advance for not naming individual contributors, and thank instead the whole CIVITAI community.

I developed it for fast response and LORA compatibility, and I often combine in workflows with Z image Turbo, Qwen or more specialized SDXL models. It tends to work well with all sorts of SDXL, Illustrious and PNY LORAs, and outputs are good at 1280x1280 resolution, but still works OK at higher ones.

The first 9 sample images are direct output from the model alone (sometimes with face refiner) without LORAs, but results can be greatly improved with a variety of LORAs and embeddings, as shown in the rest of the showcased images.

Although not created for NSFW stuff, the model’s genes include that capability and it tends to add gratuitous nudity, so be careful when prompting, being specific about clothing, appropriate subject ages, etc. so as to avoid undesirable effects.

OPTIMAL SETTINGS: The model yields rather different results depending on the sampler / cfg / noise shift combination chosen, and output will be terrible if wrongly paired. It is very fast, particularly with LCM or TCD sampling. I recommend:

CLIP Skip: 2

LCM sampler / CFG 0.9-3 / 16+ steps / noise shift 1-1.3 

TCD sampler  CFG 0.9-3 / SSS 30-35% / 16+ steps / noise shift 1.2

Euler A AYS sampler / CFG 0.9-2 / 8-16 steps/ noise shift close to 1.

DPM++ SDE Substep / CFG 0.9-1-2 / 16 steps/ noise shift 0.9-1