metadata

license: openrail++
base_model: diffusers/stable-diffusion-xl-base-1.0
tags:
  - stable-diffusion-xl
  - stable-diffusion-xl-diffusers
  - stable-diffusion
  - text-to-image
  - lora
  - loraxl
  - di.FFusion.ai
language:
  - en
thumbnail: >-
  https://cdn-uploads.huggingface.co/production/uploads/6380cf05f496d57325c12194/p54u7dEP1u8en0--NMEjS.png
library_name: diffusers

FFusion LoRA Extracted Models for Hugging Face 🚀

Welcome to the FFusion LoRA extracted models repository on Hugging Face! Here, we present a collection of models extracted using the Low-Rank Adaptation (LoRA) technique to provide a rich dataset for research and further exploration.

📦 Base Models

The following models served as the foundation for our extractions:

🔍 Extraction Details

Variants: Each base model was extracted into 4-5 distinct variants.
Extraction Depth: The models uploaded here contain approximately 70% of extracted data. These extractions yield a dataset size of around 400 GB.
Precision: We experimented with both float32 and float64 for optimal extraction results.
Differences Measurement: Singular Value Decomposition (SVD) was utilized to measure differences between the original and the tuned models. A threshold of (1 \times 10^{-3}) was commonly used, although in some cases, (1 \times 10^{-5}) and (1 \times 10^{-2}) were tested.
Demonstration Parameters: For our demonstration, we employed "conv_dim": 256 and "conv_alpha": 256.

⚙️ Technical Notes

Most SDXL models in this collection are not traditionally "trained." Instead, they are merged from previous SDXL 0.9 versions or created using other methods with the help of Comfy UI.
An important note for users: all models saved with Comfy add an extra key text_model.encoder.text_model.embeddings.position_ids. We've made necessary adjustments to ensure compatibility with the current scripts from Kohoya.

📈 Use Cases

These extracted models are intended for research and testing. They can be particularly useful for:

Investigating the potential of merging multiple LoRAs.
Weighting experiments with 1-5 LoRAs simultaneously.
Exploring the differences and similarities between LoRAs extracted from different base models.

🎉 Acknowledgements & Citations

A huge shoutout to the community for their continued support and feedback. Together, we are pushing the boundaries of what's possible with machine learning!

We would also like to acknowledge and give credit to the following projects and authors:

ComfyUI: We've used and modified portions of ComfyUI for our work.
kohya-ss/sd-scripts and bmaltais: Our work also incorporates modifications from kohya-ss/sd-scripts.
lora-inspector: We've benefited from the lora-inspector project.
KohakuBlueleaf: Special mention to KohakuBlueleaf for their invaluable contributions.

FFusion LoRA Extracted Models - How to Use Guide 🧠

Welcome to the technical guide for using the FFusion LoRA extracted models. This document will walk you through the steps required to fuse LoRA parameters, load checkpoints, and perform inference.

Fusing LoRA Parameters 🔗

To merge the LoRA parameters with the original parameters of the underlying model(s), leading to a potential speedup in inference latency:

pipe.fuse_lora()

Unfusing LoRA Parameters ⛓️

To revert the effects of fuse_lora():

pipe.unfuse_lora()

Working with Different LoRA Scales 🎚️

To control the influence of the LoRA parameters on the outputs:

pipe.fuse_lora(lora_scale=0.5)

Working with FFusion Models 🔍

Here's how to load and utilize our FFusion models:

from diffusers import DiffusionPipeline
import torch

pipeline_id = "FFusion/FFusionXL-BASE"

pipe = DiffusionPipeline.from_pretrained(pipeline_id, torch_dtype=torch.float16)
pipe.enable_model_cpu_offload()

lora_model_id = "FFusion/400GB-LoraX"
lora_filename = "FFai.0038.Realitycheckxl_Alpha11.lora.safetensors"
pipe.load_lora_weights(lora_model_id, weight_name=lora_filename)

prompt = "papercut sonic"
image = pipe(prompt=prompt, num_inference_steps=20, generator=torch.manual_seed(0)).images[0]
image

Running Inference 🖼️

After loading the desired model, you can perform inference as follows:

generator = torch.manual_seed(0)
images_fusion = pipe(
    "masterpiece, best quality, mountain", output_type="np", generator=generator, num_inference_steps=2
).images

Library of Available LoRA Models 📚

You can choose any of the models from our repository on Hugging Face or the upcoming repository on CivitAI. Here's a list of available models with lora_model_id = "FFusion/400GB-LoraXL":

lora_filename = 
    - FFai.0001.4Guofeng4xl_V1125d.lora_Dim64.safetensors
    - FFai.0002.4Guofeng4xl_V1125d.lora_Dim8.safetensors
    - FFai.0003.4Guofeng4xl_V1125d.loraa.safetensors
    - FFai.0004.Ambiencesdxl_A1.lora.safetensors
    - FFai.0005.Ambiencesdxl_A1.lora_8.safetensors
    - FFai.0006.Angrasdxl10_V22.lora.safetensors
    - FFai.0007.Animaginexl_V10.lora.safetensors
    - FFai.0008.Animeartdiffusionxl_Alpha3.lora.safetensors
    - FFai.0009.Astreapixiexlanime_V16.lora.safetensors
    - FFai.0010.Bluepencilxl_V010.lora.safetensors
    - FFai.0011.Bluepencilxl_V021.lora.safetensors
    - FFai.0012.Breakdomainxl_V03d.lora.safetensors
    - FFai.0013.Canvasxl_Bfloat16v002.lora.safetensors
    - FFai.0014.Cherrypickerxl_V20.lora.safetensors
    - FFai.0015.Copaxtimelessxlsdxl1_V44.lora.safetensors
    - FFai.0016.Counterfeitxl-Ffusionai-Alpha-Vae.lora.safetensors
    - FFai.0017.Counterfeitxl_V10.lora.safetensors
    - FFai.0018.Crystalclearxl_Ccxl.lora.safetensors
    - FFai.0019.Deepbluexl_V006.lora.safetensors
    - FFai.0020.Dream-Ffusion-Shaper.lora.safetensors
    - FFai.0021.Dreamshaperxl10_Alpha2xl10.lora.safetensors
    - FFai.0022.Duchaitenaiartsdxl_V10.lora.safetensors
    - FFai.0023.Dynavisionxlallinonestylized_Beta0371bakedvae.lora.safetensors
    - FFai.0024.Dynavisionxlallinonestylized_Beta0411bakedvae.lora.safetensors
    - FFai.0025.Fantasticcharacters_V55.lora.safetensors
    - FFai.0026.Fenrisxl_V55.lora.safetensors
    - FFai.0027.Fudukimix_V10.lora.safetensors
    - FFai.0028.Infinianimexl_V16.lora.safetensors
    - FFai.0029.Juggernautxl_Version1.lora_1.safetensors
    - FFai.0030.Lahmysterioussdxl_V330.lora.safetensors
    - FFai.0031.Mbbxlultimate_V10rc.lora.safetensors
    - FFai.0032.Miamodelsfwnsfwsdxl_V30.lora.safetensors
    - FFai.0033.Morphxl_V10.lora.safetensors
    - FFai.0034.Nightvisionxlphotorealisticportrait_Beta0681bakedvae.lora_1.safetensors
    - FFai.0035.Osorubeshialphaxl_Z.lora.safetensors
    - FFai.0036.Physiogenxl_V04.lora.safetensors
    - FFai.0037.Protovisionxlhighfidelity3d_Beta0520bakedvae.lora.safetensors
    - FFai.0038.Realitycheckxl_Alpha11.lora.safetensors
    - FFai.0039.Realmixxl_V10.lora.safetensors
    - FFai.0040.Reproductionsdxl_V31.lora.safetensors
    - FFai.0041.Rundiffusionxl_Beta.lora.safetensors
    - FFai.0042.Samaritan3dcartoon_V40sdxl.lora.safetensors
    - FFai.0043.Sdvn6realxl_Detailface.lora.safetensors
    - FFai.0044.Sdvn7realartxl_Beta2.lora.safetensors
    - FFai.0045.Sdxl10arienmixxlasian_V10.lora.safetensors
    - FFai.0046.Sdxlbasensfwfaces_Sdxlnsfwfaces03.lora.safetensors
    - FFai.0047.Sdxlfaetastic_V10.lora.safetensors
    - FFai.0048.Sdxlfixedvaefp16remove_Basefxiedvaev2fp16.lora.safetensors
    - FFai.0049.Sdxlnijiv4_Sdxlnijiv4.lora.safetensors
    - FFai.0050.Sdxlronghua_V11.lora.safetensors
    - FFai.0051.Sdxlunstablediffusers_V5unchainedslayer.lora.safetensors
    - FFai.0052.Sdxlyamersanimeultra_Yamersanimev2.lora.safetensors
    - FFai.0053.Shikianimexl_V10.lora.safetensors
    - FFai.0054.Spectrumblendx_V10.lora.safetensors
    - FFai.0055.Stablediffusionxl_V30.lora.safetensors
    - FFai.0056.Talmendoxlsdxl_V11beta.lora.safetensors
    - FFai.0057.Wizard_V10.lora.safetensors
    - FFai.0058.Wyvernmix15xl_Xlv11.lora.safetensors
    - FFai.0059.Xl13asmodeussfwnsfw_V17bakedvae.lora.safetensors
    - FFai.0060.Xl3experimentalsd10xl_V10.lora.safetensors
    - FFai.0061.Xl6hephaistossd10xlsfw_V21bakedvaefp16fix.lora.safetensors
    - FFai.0062.Xlperfectdesign_V2ultimateartwork.lora.safetensors
    - FFai.0063.Xlyamersrealistic_V3.lora.safetensors
    - FFai.0064.Xxmix9realisticsdxl_Testv20.lora.safetensors
    - FFai.0065.Zavychromaxl_B2.lora.safetensors

📊 Text Encoder Difference Overview Based on the extraction process, we observed the following differences in the text encoder across various models: