3xThreat
Full FP32 Model
Tested with SDXL, PONY and illustrious Lora's
Custom CLIP
Works with many high quality trained loras - such as Princess PONY, illustrious Skin Realism, SDXL Lightning (LORA's that relied heavily on model likeness may not work)
I have been testing MakeThemComeAliveAIArt character models with AIDigitalMediaAgency realistic skin or mid -journey models with good success
All examples use COMFY UI cmd /k python main.py --fp32-text-enc --fp32-vae --bf16-unet
Normal/Euler A with a 2x upscale and SGM_Uniform/Dies or ipndm (Img to Img) has become my go to.
The better you describe the scene and details, the higher quality the results.
Description
FAQ
Comments (45)
Wow, this is wild, Fell. I use all of your clips and checkpoints, but this is really interesting! Is the custom clip baked in?
Everything is baked in at FP32
@Felldude Great! Thanks for pushing the boundaries once again! <tips fedora>
great what frackin folder 2 put it....lol.. ty
droppin it in root.. lol.. messin up my organizing with this anomaly X } thank you for making something different
It loads as a standard checkpoint with clip and vae baked in
Right? I also dropped it in root. I'm going to have to make a felldude folder.
@Felldude lol... i know i am organized sdxl/ pony/ illus/ flux etc.. all get separate folders.. i was joking this is three of those .. how would i choose...
@mystifying I have that for my LORA's but not my checkpoints, my couple of TB are close to full as it is lol
@mystifying I have checkpoints and then loras organized that way too and then loras subfolder'd by category- concept, style, clothes, characters, celebrity, backgrounds etc. like 12-14 folders or so (using civitai's organizational naming schemes)
I love this model. It's versatile, adheres to prompts nicely and works really well with Pony and SDXL loras! (I haven't tested Illustious though). Some loras seem to over bake the image a bit but dialing them down works.
TIP- In Comfyui, if it doesn't understand a concept, location, object, you can use the "model merge" node and inject a little SDXL (or Pony) into it, they play nice together! ;)
Great job on this one
Thank you!
Thank you
This model is amazing, I never seen a model for that turns unrealistic character LoRA's into realistic beings as good as this
Thanks
Next Felldude model will be a FP64 SD1.5 checkpoint with 2 VAE and 4 T5 CLIPS.
It will beat flux.
More seriously (if one can say so) I took you at your word and launched prompts with several wildcars including Pony and Illustrious Lora.
Apart the backgrounds which remained very monotonous (maybe because of loras), I didn't have many burned images (but given the stupid random prompt, they are not interesting).
Float 64 support was added in Comfy, presumably for Hunyaun video timestamps
I think there might be some fp32 SD1.5 checkpoints. Not sure it would have any benefit.
And you can have T5 XL with SD1.5 .. just wont work with much hardcore NSFW stuff, due.. well T5.
@Mescalamba SDXL was trained by salesforce at FP32 also, the T5 was trained by google at FP32 - but to my knowledge no large model exist trained at FP64 or 128, 256 arch
@Felldude I didn't tell that seriously, I just thought it's funny how people are making "SDXL better than SDXL".
About 64bits models, I think the ratio of gain or quality (if there is any!) vs.performance would be too low.
@Brewce I think a CLIP model trained in FP64 would improve all models, but it would take a large CPU farm that wasn't being used to feed GPU's
I tried it and no good results. i get low quality images. Using forge ui. ME got_ Utdated knowledge. Do i have to run this withcomfy ui?? There is no explanation
Forge still has some issues with its GPU slider and FP32 models - Comfy would be my recommendation using the launch: cmd /k python main.py --fp32-text-enc --fp32-vae --bf16-unet
@Felldude Hi thank you for the reply. I guess i need to learn some new tricks. Ty for replying and ty for trying to break new ground!
I'm glad FP32 models are starting to take off. Great work on this one!
Thank you
how to make sfw? this wants to make nsfw
I use no humans positive then 1girl, female, nude, in the negatives
@Felldude that helped but still leaning towards hotness.. at least with the current challenge lora.. scribbleray.. will try it with next challenge
@mystifying Im sure its the tons and tons of nsfw training the model has.
@Felldude lol.. figured which is good for majority of people here... for pg completion.. not so much..lol however i will come back to each time to test it with the lora of the challenge
@mystifying Try rating_safe at the beginning of the positive prompt and nudity at the beginning of the negative. Using rating_explicit and/or rating_questionable in the negative might help as well, but I haven't needed to try that with this model yet.
@AFD_0 ty 4 extra info
"this wants to make nsfw"
You had me at 'this'...
https://civitai.com/images/59816385
i used all 3 sdxl,pony, illus i think to many loras knocked it out of adding to page.
Having all lora's work with 3 highly divergent models may not be possible, but I do know it works with SDXL lightning, a skin realism lora for illustrious, and most pony character lora's I tested
All Pony model can use XL and Illustrious lora, which are NOT character and vice versa.
Question is, can it use celeb, character loras of all 3 models? Or more important, has it likeness?
The CLIP-L for PONY, illustrious and SDXL, is completely different. That has been altered for the model, and the UNET with it.
how the hell someone even trains something like this? The only method I can think of is just training the model while injecting different loras into it to keep the same noise as it did before
diffusion, and the law of large numbers - that and the fact all these models use the same VAE, and very minor changes to the CLIP
Prompt is natural language, booru tags, both?
both
nice
The hands on this model are exceptional.
Thank you
I loaded it same as other normal 6Gb Pony checkpoints, with same parameters - it works the same as others, no issues. (not using start up arguments, ComfyUI).
Prompts understanding seems good, it renders almost exactly same pose as prompted image, (I used w14 tagger to make tags from photos).
After some tests I noticed that hands are kinda messed, with recommended f32 arguments, it improved hands a little, but not ideal.
Awesome model!
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.