Finally a model merge I got exciting enough for to post here after a long time of lurking and trying out different models, prompt styles and getting to know AI. I've always been a big fan of realism and prompting for good quality images can be difficult and delicate. That's why I wanted to work on a model that is easy to use, produces decent to high quality realistic images under the lightweight engine of SD1.5 and fp16. Thank you for reading, without further ado, here is the model.

Insane Deliberate Realism

I chose this name as I based most of my work on InsaneRealismV2.0 and PerfectDeliberate. Other models and trainings have been added in various ratios until I was satisfied with the result.

This model can handle both simple and complex prompting, with or without negatives, does reasonably well with LORAs and can handle textual inversion embeddings just fine. It seems to handle ethnicity in a (for me) acceptable way, though it sometimes tends to have a default face if the prompt is really simple. I can't wait to see what the expert prompt engineers (if you all are still working with SD1.5) can do with this!

For ease of use, a VAE is baked in, which is FtMse840000.vae.pt, but feel free to use your own.

Pointers and Usage

The model will work fine with just about any prompt, though it sometimes leans towards NSFW. In that case simply add (NSFW:1.1) at the start of your negative prompt.

You don't need quality prompt tokens like (masterpiece) or (high resolution) as you will see in the examples, but this model seems to enjoy the following:

BREAK (realistic), (incredible detail), real skin, high resolution, 16K, HDR, grainy, Fujifilm XT3, 115mm lens

You can add it at the end of your prompt or save it as a style option if you're using A1111. Similarly, the model doesn't really need a negative prompt, but if you want a default that works well, you can use this:

(worst quality, low quality:1.4), watermark, signature, username, artist name, artist tag, copyright, website, logo, monochrome, extra limb, wrong anatomy, disfigured, (mutated hands and fingers)

I will include a picture with the same prompt with and without these "enhancers" so you can see the difference clearly.

Resolutions can be anything, though it likes rectangles better at higher resolutions. Examples were made with the following settings:

Sampler: Euler a
Sampling steps: 50
Hires.fix with upscaler R-ESRGAN 4x+, upscale by 2, Hires steps 30, denoising 0.45
Resolution: 600 x 960
CFG scale: 7
Clipskip: 1

Clipskip 2 also works if the prompt is a bit unruly, aDetailer is optional so use at your own discretion.

Lastly, the model is aimed towards realism and it's pretty hard to make it do anything else. Even using keywords like (anime style) can be used without it generating actual anime or drawing like results, though I haven't tested this in very complex prompts.