veraPrime is a v-pred model merging models in the Noob v-pred lineage.
30 Steps, 5-7 CFG
EulerDiscreteScheduler.from_config({
pipe.scheduler.config,
{
"prediction_type": "v_prediction",
"rescale_betas_zero_snr": True
}
})Description
FAQ
Comments (13)
What are the settings/parameters for generation with this? No matter what settings I use, it comes out all pixels and goop :D
Also, if this is vpred, doesn't it need a config file? :)
@DarkArtsAI I run this with diffusers and use:
EulerDiscreteScheduler.from_config({ p.scheduler.config, {"prediction_type": "v_prediction", "rescale_betas_zero_snr": True }})
and it works great
Also, I'm just curious - why do you expect an additional config file?
@epiTune Because google XD I'm trying to use this with Fooocus, and I read somewhere that vpred models most of the time come with a config file. Is there a way to run this with either Fooocus or Forge UI?
@DarkArtsAI I'm not aware if Fooocus supports v-pred - a quick search online doesn't turn up any clear evidence that they do, but didn't look that much.
But yes you would have to use a tool that supports v-pred to use this model.
The Noob model card states that reForge has support apparently FWIW
@epiTune Ok, cheers! I'll do some research then and find a tool that I can use it with :3
Please add v_pred and ztsnr keys. This model is not detected as a v_pred model because the required keys are missing from the model's metadata; one needds to manually activate Advanced Sampling in order to use it. This can be easily avoided if the keys are added to metadata. The original script is published at https://civitai.com/models/1027203/noobvpencil-xl (check the "About this model" section).
Add v_pred and ztsnr keys:
# Python Script BASE_MODEL_NAME = "noob_v_pencil-XL-v0.5.1" tensors = {} with safe_open(BASE_MODEL_NAME + "-base.safetensors", framework="pt", device="cpu") as f: for key in f.keys(): tensors[key] = f.get_tensor(key) tensors["v_pred"] = torch.tensor([0.0]) tensors["ztsnr"] = torch.tensor([0.0]) save_file(tensors, BASE_MODEL_NAME + ".safetensors")Idk - that seems like way too much of a hack
It is not. It's just a proper way to mark the model as v_pred in its metadata, which allows tools such as reForge to work properly (this model only works with a manual override). Please check the "About this model" section for the linked "noobvpencil-xl" model, where the author explains it better than I can.
@civit77899 Yeah I get that the tools have these guessing heuristics implemented, for example:
https://github.com/lllyasviel/stable-diffusion-webui-forge/blob/f53307881bfd824dbdce6ac0d4bba04d9a74ab36/backend/loader.py#L249
But what I'm saying is that's it's FOR SURE a hack to add dummy vectors to the safetensors just for something to exists that allows these checks.
So it's not a proper way - it's just a hack that works in multiple UIs.
I will still add it I suppose - but I wish something less hacky and more well defined existed.