dark_cinematic_cute v2.2 - Illustrious Redux

tldr;

Positive Prefix: "2.5D, semi-realistic, cinematic, details, BREAK"

Negative Prefix: "text, logo, watermark, bad hands, bad feet"

Steps: 32-40

CFG: 7-8

Sampler: euler a (fastest), dpmpp_2m (good), gradient estimation (best)

Scheduler: normal, karras (good), beta (best)

Resolutions: SDXL resoluions, 1280x1280

purpose

This model is designed for semi-realistic, cinematic shaded, pornography with a focus on female subjects. This merge is based on DCCv2.0, but has a different style with greater focus on more realistic shading while maintaining a more artistic background. Compared to DCCv2.0, this model handles more complex scene (and sexual) descriptions better, but does not handle furry/anthro as well. If your goal is furry/anthro, stick with v2.0 or v2.1 if wanting to try extreme resolutions.

dark_cinematic_cute v2.1 - Illustrious HD

tldr;

Positive Prefix: "2.5D, semi-realistic, cinematic, BREAK"

Negative Prefix: "text, logo, watermark, bad hands, bad feet"

Steps: 32-40 -=- CFG: 4-6

Sampler: euler a (fastest), dpmpp_2m (good), deis (best)

Scheduler: normal, karras (good), beta (best)

Resolutions: 1536x1536, 1824x1248, 1248x1824

purpose

This is an experimental version of DCCv2.0, which has been merged with a version of Illustrious v1.0 that was trained on using dreambooth to improve 2.5D performance. The result is a model which is nearly as good as DCCv2.0 with the largest difference being less detailed backgrounds, but with the benefit of consistently supporting resolutions up to 1536x1536. It does a very good job for non-realistic porn in general but is very myopically designed to do a specific style of 2.5d semi-realistic porn with a heavy slant toward a cinematic shading style. This model is also designed with semi-anthro (e.g. nekomimi) in mind and also handles furry with a female subject well.

dark_cinematic_cute v2.0 - Illustrious

tldr;

Positive Prefix: "2.5D, semi-realistic, cinematic, BREAK"

Negative Prefix: "text, logo, watermark, bad hands, bad feet"

Steps: 32-40

CFG: 4-6

Sampler: euler a (fastest), dpmpp_2m (good), deis (best & slow)

Scheduler: normal, karras (good), beta (best)

Resolutions: SDXL resoluions, 1280x1280

purpose

Been doing merges for my personal use for quite a while, and this is my fourth model made specifically for public consumption. It does a very good job for non-realistic porn in general but is very myopically designed to do a specific style of 2.5d semi-realistic porn with a heavy slant toward a cinematic shading style. This model is also designed with semi-anthro (e.g. nekomimi) in mind and also handles furry with a female subject well.

dark_cinematic_cute v1.0 - PDXL

tldr;

Positive Prefix: "score_9, score_8_up, score_7_up, BREAK, 2.5D, semi-realistic, cinematic, BREAK"

Negative Prefix: "score_4, score_5, score_6, BREAK, text, logo, watermark, bad hands, bad feet"

Steps: 28-42

CFG: 7-9

Sampler: euler a (fastest), dpmpp_2m (preferred), heun (best & slow)

Scheduler: normal, karras (preferred), sgm uniform, beta

Resolutions: 1024x1024, 1216x832, 832x1216

Note: CFGs of 10+ are attainable with steps in the 80+ range, but quality increase faces extreme diminishing returns.

purpose

Been doing merges for my personal use for quite a while, and this is my second model made specifically for public consumption. It does an acceptable job for non-realistic porn in general but is very myopically designed to do a specific style of 2.5d semi-realistic porn with a heavy slant toward a cinematic shading style. By the merit of that exclusive focus, generation in all other styles is significantly inferior to other PonyXL models. For that reason, tags such as "realistic," "photograph," or "anime" will degrade quality. This model is a merge but also incorporates a LoRA-based improvement process I developed for my own use, which is detailed below.

merge

This is a merge of waiCUTE v6.0, SEMImergeij v5.0, Hs UltraHD CG Pony Ver., and CuteCandyMix Pony Ver, and two LoRAs, both of which I trained myself. I could be willing to upload those two LoRAs if there is interest, but essentially, they are just a tool for altering the style and encouraging a few design elements I am personally a fan of.

lora process

The primary differentiator of this model is that it comes prebaked with LoRAs trained on images generated by an earlier versions of itself. This greatly harms the versatility of the model but allows for much more strict adherence to style guidelines and lets me reinforce a few traits I like. The process is as follows: First, generate about a thousand images using a custom ComfyUI plugin to generate random prompts, saving a cut-down version of the prompts to a txt file of the same name. Next, remove images that are poor quality or break from defined style guidelines. Next, select a few hundred of the best images and train a LoRA on them. Repeat the process a couple of times. Next, create sets of images with a control and a few different variations of LoRAs and LoRA weights. Next, conduct a blind comparison between sets of images, awarding points to adherent images. The final merge model - LoRA combination is the one that consistently wins. Is there a better way to do this? A faster, less tryhard way? Almost definitely, but this is more fun, and I'm doing it for free. The prompt generator I wrote for this task is called PromptByMood and can be found at: https://github.com/RLW-Chars/comfyui-promptbymood