John6666/DiffuseCraftMod · New models and LoRAs

Oh, man, that sucks. Hopefully they'll up the limit soon. On that note, does this space require that specifically you upload it, or can you use any model that anyone uploads to Hugging Face?

John6666

Owner 5 days ago

or can you use any model that anyone uploads to Hugging Face?

this.

funnycat1

about 8 hours ago

or can you use any model that anyone uploads to Hugging Face?

this.

So, if I uploaded the model myself, would you be willing to add it to your space?

John6666

Owner about 8 hours ago

if I uploaded the model myself, would you be willing to add it to your space?

Yeah. But one thing to note: Only the model itself can be added to the model list. Adding LoRAs that way is a bit tricky (due to the structure of my space...).

funnycat1

about 8 hours ago

•

edited about 8 hours ago

Yeah. But one thing to note: Only the model itself can be added to the model list. Adding LoRAs that way is a bit tricky (due to the structure of my space...).

Nice, thanks. As far as the LoRAs, could you delete one that I've previously requested to add another? I'd like to swap DoubleExposure IL for (PTD) Style: Contrast, Glow & Lighting Enhancement.

John6666

Owner about 8 hours ago

Sorry. Currently, I'm not really in the mood to manage LoRA files either...

funnycat1

about 8 hours ago

Sorry. Currently, I'm not really in the mood to manage LoRA files either...

Alright, that's totally fine. :)

I'll try to upload some models and link them here once I'm done.

funnycat1

about 6 hours ago

Hm, I thought I could just upload the SafeTensor file from Civitai, but there seems to be more to it than that. I'm not really sure what to do.

John6666

Owner about 5 hours ago

I could just upload the SafeTensor file from Civitai,

Actually, with DiffuseCraft/DiffuseCraftMod, if it's uploaded to the HF Hub, you can use the standalone safetensors file as-is. The URL specification might be a bit confusing though...
Like this: https://huggingface.co/Raelina/Raehoshi-illust-XL-9/blob/main/raehoshi-illust-xl-9.safetensors

but there seems to be more to it than that.

Yeah. In my case, to work around bugs in the SDXL standard VAE or handle scenarios like single safetensors files without a built-in VAE, I apply conversion like the following to select the VAE, scheduler, or CLIP:

The easiest way

Use Diffusers like this:

load the single .safetensors file with StableDiffusionXLPipeline.from_single_file(...)
replace the scheduler with Euler A
replace the VAE with madebyollin/sdxl-vae-fp16-fix
save the pipeline with save_pretrained(...)

This works because Diffusers supports loading SDXL from a single file, and save_pretrained() writes the normal Diffusers folder structure with subfolders. In Diffusers, A1111 “Euler a” = EulerAncestralDiscreteScheduler. The sdxl-vae-fp16-fix VAE was made to avoid SDXL fp16 NaNs, and its model card says it also works with SDXL 1.0. (Hugging Face)

Install once

Install PyTorch.
Install Diffusers

pip install -U diffusers transformers accelerate safetensors

Diffusers’ install docs list diffusers and transformers, and the model-format docs note that safetensors support requires the Safetensors library to be installed. (Hugging Face)

One-model converter script

Save this as convert_one_sdxl.py.

import os
import gc
import torch

from diffusers import (
    StableDiffusionXLPipeline,
    AutoencoderKL,
    EulerAncestralDiscreteScheduler,
)

# ==== EDIT THESE 2 LINES ====
input_file = r"/path/to/your_model.safetensors"
output_dir = r"/path/to/your_model_diffusers"
# ============================

print("Loading SDXL safetensors file...")

pipe = StableDiffusionXLPipeline.from_single_file(
    input_file,
    use_safetensors=True,
    torch_dtype=torch.float16,
)

print("Setting scheduler to Euler A...")
pipe.scheduler = EulerAncestralDiscreteScheduler.from_config(pipe.scheduler.config)

print("Loading fp16-fix VAE...")
pipe.vae = AutoencoderKL.from_pretrained(
    "madebyollin/sdxl-vae-fp16-fix",
    torch_dtype=torch.float16,
)

print("Saving as Diffusers folder...")
pipe.save_pretrained(
    output_dir,
    safe_serialization=True,
)

del pipe
gc.collect()

if torch.cuda.is_available():
    torch.cuda.empty_cache()

print("Done.")
print(f"Saved to: {output_dir}")

This script always uses torch.float16, saves the pipeline in Diffusers folder format, and makes the saved folder default to Euler A and fp16-fix VAE because those components are set on the pipeline before saving. save_pretrained() is the Diffusers-supported way to create the folder layout. (Hugging Face)

Run it

python convert_one_sdxl.py

After it finishes, output_dir will contain a normal Diffusers model folder instead of one big checkpoint file. (Hugging Face)

What each important line does

from_single_file(...)
Loads your single SDXL .safetensors checkpoint into Diffusers. (Hugging Face)
torch_dtype=torch.float16
Forces the pipeline to load in fp16. (Hugging Face)
EulerAncestralDiscreteScheduler.from_config(...)
Changes the scheduler to Euler A. (Hugging Face)
AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", ...)
Replaces the VAE with the fp16-safe SDXL VAE fix. (Hugging Face)
save_pretrained(..., safe_serialization=True)
Writes the Diffusers folders and saves with safetensors serialization. (Hugging Face)

Minimal answer

Use this pattern:

pipe = StableDiffusionXLPipeline.from_single_file(
    input_file,
    use_safetensors=True,
    torch_dtype=torch.float16,
)

pipe.scheduler = EulerAncestralDiscreteScheduler.from_config(pipe.scheduler.config)

pipe.vae = AutoencoderKL.from_pretrained(
    "madebyollin/sdxl-vae-fp16-fix",
    torch_dtype=torch.float16,
)

pipe.save_pretrained(output_dir, safe_serialization=True)

That converts one SDXL 1.0 .safetensors file into a Diffusers folder, with defaults set to Euler A, fp16-fix VAE, and fp16 loading in the converter. (Hugging Face)