This is a collection of popular Pony models transformed to FP8.
Advantages:
Half the size
Same Quality
Less VRAM required
Description
FAQ
Comments (14)
Is there some magic wand to wave to get the lower VRAM usage? I'm still getting the same usage when comparing 1:1 with the original models.
in Forge the original pony used 16gb and the FP8 version only 14gb. Its not as drastic as the checkpoint size reduction but its still about 20% less.
This is great. Will you be converting other FP16 models into FP8. How can it be done in ComfyUI?
I was planning to, do you want a specific one?
I don't use Comfy UI so I can not tell.
@JohnnyB1 How do you convert a Pony model to FP8. Which program do you use? Can you share your steps? Thanks.
its pretty easy, all you have to do is take a fp8 model and merge another model with very high percentage (0.99 or even 1) into it. I do it with super merger in auto1111.
@JohnnyB1 Hi, hope you don't mind me piggybacking off this but as for 'want a specific one', could I throw out the suggestion for "Js2Prony"? It's also one of those older but long lasting ones that could do with saving space.
@cookiecat I forgot about your request, JS2prony coming right up any minute now :)
@cookiecat I learned how to make any Pony (PDXL) and SDXL checkpoint into FP8. The results are incredible as they have nearly the same quality of out - especially if you're going to upscale and use adetailer. They use half the vram too.
How did you quantize it?
The easiest way is just to install "ComfyUI_DiffusionModel_fp8_converter" node in ComfyUI.
Put "Load checkpoint", "Save checkpoint" nodes and put "ModelFP8ConverterNode" between them. It will save quantized model in .safetensors file.
Any more models planned? Or have you wrapped up this series
the response wasn't what I hoped for, and also the tool I used was not working for some time after an update. So currently I have no plans, do you have a request?
@JohnnyB1 It may be worth your time investing in FP8 versions of certain Illustrious models, such as KiwiMix or ObsessionIllustrious. With Illustrious blowing up at the moment, I can see people wanting quantized versions of it.
As for pony models, I'd like to see FP8 versions of models such as PrefectPony (popular), CuteCandyMix, t-PonyNAI, and bb271 (a favorite). All of these models produce different styles and various different things. Another model worth mentioning is Hadrian DeliceXL, but that's a lesser popular one and produces a lot of the same styles from models I'd already mentioned.
I find these FP8 models amazing. They're great for doing other things while testing LoRAs on smaller cards like the one on my laptop. I will work on a bunch of examples for the Pony & Autism versions.