illustrious ULtraReal (6GB)
This mixed precision model handled SFW and NSFW prompts extremely well in testing.
Only 6GB
Can be used with the FP32 CLIP commands for added precision.
FP8 UNET had minimal losses and a had additional training to amplify detail at lower precision
This model is under 6Gb but packs a hell of a punch, I had to stop myself from feeding it your prompts as test.
Description
FAQ
Comments (40)
I've tested this model, it is perfect in terms of quality and speed!
Thanks
Easily the best realistic Illustrious checkpoint I've used. It knows less sex poses than traditional Illustrious but more than any realistic one plus the prompt comprehension is miles ahead. It also does most ethnicities well.
Thank you
@Felldude
When I tried to load the model, an error occurred midway and I couldn't load it.
safetensors_rust.SafetensorError: Error while deserializing header: InvalidHeaderDeserialization
There was an article about this error on civitai.
https://civitai.com/articles/7336/notes-on-the-error-message-invalidheaderdeserialization
Do you know if this can be resolved by changing the training settings or by changing my environment settings?
I'm using v1.7 of AUTOMATIC1111/stable-diffusion-webui.
My recommendation will be to use comfy, possibly swarm.
I would recommend FORGE or ReForge as they are evolved versions of auto1111 which is kind of sputtering along now. Very familiar interface so won't take long to get generating again. Just search Forge UI and ReForge UI to find the repos on github.
Comfy is a whole lot different but worth the time to learn too and has some features the others don't.
@fantaseed What GPU are you using? I'm using a 2080 Ti and get the exact same error message when trying to load different FP8 models (including this one). Some of the older cards don't have native FP8 support, so maybe that's it?
@Felldude @BradOh @AFD_0
It worked with comfy and reforge.
I'm using a Mac.
I've used dozens of models, but this is the first time I've seen an error that it can't be read.
@fantaseed Thanks for the update!
Can I use this model to inpainting? Im not familiar with Illustrious.
Well all I can say is WOW. It was worth waiting for this version of your model.
Easily the most realistic model I have used. Responds very well to my prompting style. Any Lora I have used with it have worked well.
Even tried some SDXL and Pony Lora. While the likeness was there it wasn't perfect for any characters I tried. Produced images didn't come out deformed.
I have some work ahead of me for training Lora just for this model.
Excellent work Felldude, can't wait to see how this model evolves.
Edit: When I tried the SDXL and Pony lora I wasn't aware of something at first. Figured it out when I watched the command console and saw this torch.float8_e4m3fn when it loading the model.
This tidbit of info is for ForgeUI users. At the top of screen set the to "All". Then set the "Diffusion in Low Bits" to "Automatic FP16 Lora."
All lora I have now experimented with have exceptional likeness.
So it begs the question "is there a version 2 in the works for this model"
Thank you
Oh my god, the hands... it seems like we're back to the days of SD 1.5.
Eventually we will just train with mittens to solve the issue, or leave the turtles, and mew out of training sets lol
This is an awesome model. It can easily turn your 2d or 2.5d render into realism so you can get all sorts of sizes and shapes normally not possible with realism!
Thank you
too much penetration of the models, I didn't ask for the male to be in there at all, and yet image after image she's mounted on him
does it persist when using the negatives source_explicit, 1boy, penis, etc
This issues on all IL checkpoints, they poorly trained initial model, tagged "spread legs" but skipped male and penis, so even negatives will not work - male is part of woman, it is unremovable.
Maybe some day those "authors" will learn how to properly tag datasets and will not use porn for everything.
I always looking for new less popular checkpoints which are capable and interesting. But this one is not the case, it struggling even with simplest prompts.
- no understanding of body types
- no understanding of anal/vaginal/toy
- very limited understanding of poses.
it feels like IL dumbed down to the level of vanilla SDXL, lost all IL features.
Here too. By using any Illustrious LORA, like Lara Croft, if I put 'curvy' or 'voluptous' and/or 'thick thighs', chubby, etc, the model remains the same (slim body). I also tried 'venus' LORAS. By changing to another checkpoint, say, hsUltrahdCG_IllEpic, it worked normally!
@Felldude your model is very good how do I change the nationality and age of the person it will have future updates could you make a version for Japanese girl psl
It would need additional training, but you could try certain illustrious trained characters to see if they match an ethnicity
Does anyone know what
"FP32 CLIP commands" means?
That would be a launch command argument. Depending on weather your using forge or comfy it is slightly different and can be added to the batch file
Ah, I googled but could not find it. In comfy, do you know the flag?
@Felldude Thank you
@Felldude Forge command line?
@lkjulpolak I use set COMMANDLINE_ARGS= --unet-in-bf16 --vae-in-fp32 --cuda-malloc --clip-in-fp32 --ckpt-dir "Your model folder if you have multiple installs"
I loved this model, its so realistic! But it doesnt respect my 'body proportion' prompts. By using any Illustrious LORA, if I put 'curvy' or 'voluptous' and/or 'thick thighs', chubby, etc, the model remains the same (slim body). I also tried 'venus' LORAS. By changing to another checkpoint, it worked normally!
I might revisit illustrious in future to look into this - I looked into it as a request and don't have the time investment like PONY or SDXL
This is why I love Pony models so much as they respect the body proportions that you
prompt compared to other base models.
Would love to see Illustrious models have that same body proportion control.
I've noticed that quite a few models, mainly the realism ones (and anything involving FLUX), seem to struggle with "same-face syndrome" where every character ends up looking like the same woman from a Victoria's Secret catalogue or something.
I guess it's not just the face with these checkpoints, huh?
Masculine prompts too?
The dataset is biased female
hey can someone give me their work flow so that i can use it in comfy UI ? I cant load the json file mentioned in the huggingface in my UI, it says version is not specified
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.