FAQ
a) Q: what is unrealspice?
A: UnrealSpice is a checkpoint i merged for my own personal use. Primarily will be focusing on achieving realism but also making it capable of understanding & generating various anime concept / fetish in realism style.
b) Q: will there be unrealspice for SDXL / illustrious or others?
A: as of current state, i will only be focusing on pony, i tried illustrious but it just doesn't perform well for me, so i will remain using pony as my main model.
c) Q: how often will new version release?
A: I will only develop new version when i am unsatisfied with current version or there is a much more superior model that prompted me to do more merging. I wouldn't be actively developing new version when there is no reason to do so. As of May 2025, latest version is V5.
Setting Up: Installing Webui
The simplest way to setup is by installing stability matrix: click here
Stability Matrix is a tools to help you setup various image generation tool / webui such as comfyui, forge, A1111, reforge, swarmui, etc with ease and simple to use ui.
For this guideline, it will be based on reforge webui as this is the main one that i use.
after installing stability matrix --> go to add packages --> install reforge webui
*note that reforge webui author had announced the development is stopped, but it is still pretty usable for sdxl era checkpoints including pony and illustrious.
**if you are using latest version of reforge webui and encountered extremely slow switching when using adetailer, switch you version to main@629b756 (this is the version i am using because with latest version of reforge webui, it take almost a minute just to switch to adetailer to continue generation.)
Setting Up: Reforge Webui
After installing reforge webui, download UnrealSpice and put it into Models --> StableDiffusion folder. Launch ReforgeWebui and it should appear in the selection at the model selection.
select Unreal Spice v5 from the dropdown list and wait for it to switch.
once done, you are ready to go for image generation.
I will be listing down additional useful extension here to help with ui functionality.
Prompt All in One: click here - this is a must have, it saved a lot of time by providing you with a lot of functionality to help with prompt generation.
State manager: click here - another must have, it allow you to load all the setting that you used to generate in the previous session with ease.
to install extension, go to extension tab --> install from url --> copy the github link to url and click install.
Basic Generation
There are 2 type of generation:
txt2img: generate from nothing and simply based on the checkpoint understanding of prompts.
img2img: generate using an existing image as reference and prompts input, easier to generate an image and could help if txt2img is not working.
For guideline, i will only share on txt2img as this is the method i use 99% of the times.
UnrealSpice Recommended settings (Reforge)
Sampling method: Euler A Comfy (works best for me) / Euler A
* DPM sampling methods also works but i find that result from Euler A is most appeasing.
Schedule Type: Simple (Recommended) / Normal
CFG Scale: 4-7, personally i use CFG scale of 6
Aspect ratio: 1024 x 1024 (square), 1152 x 896 (horizontal), 896 x 1152 (vertical)
Prompting style: follow PonyXL (danbooru format, tags format, short and simple natural languages could possible aid in generation)
To generate your first image, key in your prompts into the prompt input box and click generate.
Sample prompts for you to try:
1girl, medium shot, cowboy shot, looking at viewer, t-shirt,long pants, outdoors,garden,
reference seed: 723077577 (you should be getting the same image as above using this seed)
To improve image, you can consider to add-on quality prompts and negative prompts
Prompts:
score_9,score_8_up,score_7_up,score_6_up,score_5_up,(85mm DSLR lens,DSLR photo:0.8),4k,8k,(realistic,photorealistic:1.2),photo (medium),intricate detail,high quality,(masterpiece, best quality, ultra-detailed), (cowboy shot:0.8),clean background,minimalist background,(beautiful girl,perfect skin,realistic skin:0.9),(perfect hands,perfect eyes,realistic_eyes,symmetrical eyes,Realistic face,human face,compact face,realistic proportions:0.9) BREAK 1girl,medium shot,cowboy shot,looking at viewer,t-shirt,long pants,outdoors,garden,Negative Prompts:
score_4,score_3,score_2,score_1,monochrome,grayscale,blur,blurry,cropped image,bad proportion,bad quality,awful image,worst,signature,username,watermark,website,text,word,imperfections, BREAK bad shading,bad lighting,multiple light sources,orange lighting,moody lighting,dramatic shadows,low light,cinematic lighting,dark room,shadows,overcast,overexposure,haze,bloom,lens flare,blown out lighting,overlit,excessive light,messy background,cluttered background,chaotic background,random objects,background distractions,unwanted items,overlapping props, BREAK bad hand,too many fingers,bad fingers,six fingers,missing limb,lazy eyes,bad eyes,uneven eyes,blanked eyes,tan line,tanned body,skin pores,imperfect skin,skin spot,sun spots,age spots,freckles,shiny face,glossy face,elongated face,(long face:1.3),distorted proportions,deformed face,stretched face,(elongated jaw:1.2),(face close-up:1.2),(face focus:1.2),(close-up:1.2),messy background,cluttered background,chaotic background,random objects,background distractions,unwanted items,overlapping props,
Result using the same seed.
* try to learn and use adetailer extension if your generation involves character, it can massively help to improve on the details.
For further enhancement, utilize lora, embeddings as well as webui extension (these will be shared in further guidelines), this is my result using my usual setup:
Troubleshooting
sometimes when generating images, you will get an image that is purely noise like below:
This is due to the checkpoints are unable to understand the prompts. Essentially a failed generation.
Solutions:
a) try rearranging prompt in different manners or adding more relevant prompts (i find that for me most of the time this issues is caused by prompts arrangement (prompts token count), therefore by rearranging / adding more prompts, there is a chance to fix and generate the image.
b) the checkpoints simply are unable to understand the prompts combination, try searching for a lora that matches the concept / imagery you wished to generate and add into the prompts.
c) Use another checkpoints to attempt the same generation.
The reason that i created UnrealSpice is to minimize this issue in my generation as i faced with this failed generation many time using other checkpoints. However, this does not mean that UnrealSpice will not have this issue, so if nothing above works, try finding another checkpoints you prefer to attempt the generation.
I cannot generate the image in my mind even after a lot of attempts.
a) checkpoints have it's limitation, as different checkpoints have their own core idea and prompt understanding capabilities. For example, if you try using this checkpoint (Unrealspice) to attempt generate anime / 2d image, you will be wasting your time as I created UnrealSpice for generating realism image.
b) Lora is designed to solve this issue to some extend, therefore if the checkpoints itself cannot understand the concepts, try to find if any creators in civitai had created lora for the exact concept and apply the lora to see if able to generate your expected outcome.
Conclusion
Before ending this basic guideline, i'd like to mention Image Generation is very RNG-dependant, especially with txt2img method, sometimes you will need many attempts to generate the images that meet your expectation. you can learn a lots of generation tips & tricks through the web or even by asking AI assistance such as ChatGPT to aid in your generation. I would say to improve on your image generation, the first step is to understand what are the effective prompts for your image. Easiest way to learn on prompting is by referring other's generation in civitai or other similar website and utilize their prompts. Therefore, everyone can always share your masterpiece and let others learn as well.