<a target="_blank" rel="ugc" href="https://huggingface.co/spaces/FFusion/FFusionXL-SDXL-DEMO"><img src="https://cdn-uploads.huggingface.co/production/uploads/6380cf05f496d57325c12194/STGBu43Fv8ket-sXH6JA-.png" alt="ffusionXLs.png" /></a><a target="_blank" rel="ugc" href="https://huggingface.co/FFusion/FFusionXL-09-SDXL/blob/main/FFusionXL-09-SDXL.safetensors"><img src="https://img.shields.io/badge/-Download%20Model-brightgreen?style=for-the-badge&amp;logo=appveyor" alt="Download Model" /></a><a target="_blank" rel="ugc" href="https://huggingface.co/spaces/FFusion/FFusionXL-SDXL-DEMO"><img src="https://img.shields.io/badge/-FFusionXL%20DEMO-brightpurple?style=for-the-badge&amp;logo=appveyor" alt="FFusionXL SDXL DEMO" /></a><h1 id="heading-19">FusionXL SDXL model + Diffusers</h1><pre><code>from diffusers import DiffusionPipeline
import torch

pipe = DiffusionPipeline.from_pretrained("FFusion/FFusionXL-09-SDXL", torch_dtype=torch.float16, use_safetensors=True, variant="fp16")
</code></pre><a target="_blank" rel="ugc" href="https://huggingface.co/FFusion/FFusionXL-09-SDXL"><img src="https://img.shields.io/badge/FFusionXL--09--SDXL-Model-pink" alt="FFusionXL-09-SDXL" /></a><img src="https://img.shields.io/badge/%F0%9F%94%A5%20Refiner%20Compatible-Yes-success" /><img src="https://img.shields.io/badge/%F0%9F%92%BB%20CLIP--ViT%2FG%20and%20CLIP--ViT%2FL%20tested-Yes-success" /><img src="https://img.shields.io/badge/%F0%9F%A7%A8%20FFXL%20Diffusers-available-brightgreen" /> <a target="_blank" rel="ugc" href="https://cdn-uploads.huggingface.co/production/uploads/6380cf05f496d57325c12194/iM_2uykpHRQsZgLvIjJJl.jpeg"><img src="https://cdn-uploads.huggingface.co/production/uploads/6380cf05f496d57325c12194/iM_2uykpHRQsZgLvIjJJl.jpeg" alt="ffusionXL.jpg" /></a><h2 id="heading-20">Model</h2>FFXL based on SDXL consists of a two-step pipeline for latent diffusion: First, we use a base model to generate latents of the desired output size. In the second step, we use a specialized high-resolution model and apply a technique called SDEdit (<a target="_blank" rel="ugc" href="https://arxiv.org/abs/2108.01073">https://arxiv.org/abs/2108.01073</a>, also known as "img2img") to the latents generated in the first step, using the same prompt.<a target="_blank" rel="ugc" href="https://huggingface.co/FFusion/FFusionXL-09-SDXL/blob/main/FFusionXL-09-SDXL.safetensors"><img src="https://img.shields.io/badge/-Download%20Model-brightgreen?style=for-the-badge&amp;logo=appveyor" alt="Download" /></a><h3 id="heading-21">Model Description</h3><ul><li>Trained by: FFusion AI</li><li>Model type: Diffusion-based text-to-image generative model</li><li>License: <a target="_blank" rel="ugc" href="https://huggingface.co/FFusion/FFusionXL-09-SDXL/blob/main/LICENSE.md">FFXL Research License</a></li><li>Model Description: This is a trained model based on SDXL that can be used to generate and modify images based on text prompts. It is a <a target="_blank" rel="ugc" href="https://arxiv.org/abs/2112.10752">Latent Diffusion Model</a> that uses two fixed, pretrained text encoders (<a target="_blank" rel="ugc" href="https://github.com/mlfoundations/open_clip">OpenCLIP-ViT/G</a> and <a target="_blank" rel="ugc" href="https://github.com/openai/CLIP/tree/main">CLIP-ViT/L</a>).</li><li>Resources for more information: <a target="_blank" rel="ugc" href="https://arxiv.org/abs/2307.01952">SDXL paper on arXiv</a>.</li></ul><a target="_blank" rel="ugc" href="https://cdn-uploads.huggingface.co/production/uploads/6380cf05f496d57325c12194/dtRkHom_cxGSzCV2ReeVc.png"><img src="https://cdn-uploads.huggingface.co/production/uploads/6380cf05f496d57325c12194/dtRkHom_cxGSzCV2ReeVc.png" alt="FFusionAI_00187_.png" /></a><h3 id="heading-22">Model Sources</h3><ul><li>Demo:<a target="_blank" rel="ugc" href="https://huggingface.co/spaces/FFusion/FFusionXL-SDXL-DEMO"><img src="https://img.shields.io/badge/-FFusionXL%20DEMO-brightpurple?style=for-the-badge&amp;logo=appveyor" alt="FFusionXL SDXL DEMO" /></a><a target="_blank" rel="ugc" href="https://huggingface.co/FFusion/FFusion-BaSE"><img src="https://img.shields.io/badge/Hugging%20Face-FFusion--BaSE-blue" alt="Hugging Face Model" /></a><a target="_blank" rel="ugc" href="https://github.com/1e-2"><img src="https://img.shields.io/badge/GitHub-1e--2-green" alt="GitHub" /></a><a target="_blank" rel="ugc" href="https://www.facebook.com/FFusionAI/"><img src="https://img.shields.io/badge/Facebook-FFusionAI-blue" alt="Facebook" /></a><a target="_blank" rel="ugc" href="https://civarchive.com/models/82039/ffusion-ai-sd-21"><img src="https://img.shields.io/badge/Civitai-FFusionAI-blue" alt="Civitai" /></a></li></ul><h3 id="heading-23">🧨 Diffusers</h3>Make sure to upgrade diffusers to &gt;= 0.18.0:<pre><code>pip install diffusers --upgrade
</code></pre>In addition make sure to install <code>transformers</code>, <code>safetensors</code>, <code>accelerate</code> as well as the invisible watermark:<pre><code>pip install invisible_watermark transformers accelerate safetensors
</code></pre>You can use the model then as follows<pre><code>from diffusers import DiffusionPipeline
import torch

pipe = DiffusionPipeline.from_pretrained("FFusion/FFusionXL-09-SDXL", torch_dtype=torch.float16, use_safetensors=True, variant="fp16")
pipe.to("cuda")

# if using torch &lt; 2.0
# pipe.enable_xformers_memory_efficient_attention()

prompt = "An astronaut riding a green horse"

images = pipe(prompt=prompt).images[0]
</code></pre>When using <code>torch &gt;= 2.0</code>, you can improve the inference speed by 20-30% with torch.compile. Simple wrap the unet with torch compile before running the pipeline:<pre><code>pipe.unet = torch.compile(pipe.unet, mode="reduce-overhead", fullgraph=True)
</code></pre>If you are limited by GPU VRAM, you can enable cpu offloading by calling <code>pipe.enable_model_cpu_offload</code> instead of <code>.to("cuda")</code>:<pre><code>- pipe.to("cuda")
+ pipe.enable_model_cpu_offload()
</code></pre><h2 id="heading-24">Uses</h2><a target="_blank" rel="ugc" href="https://cdn-uploads.huggingface.co/production/uploads/6380cf05f496d57325c12194/M9KPUbng7iMUlW4lV93_1.jpeg"><img src="https://cdn-uploads.huggingface.co/production/uploads/6380cf05f496d57325c12194/M9KPUbng7iMUlW4lV93_1.jpeg" alt="fusion.ai334.jpg" /></a><h3 id="heading-25">Direct Use</h3>The model is intended for research purposes only. Possible research areas and tasks include<ul><li>Generation of artworks and use in design and other artistic processes.</li><li>Applications in educational or creative tools.</li><li>Research on generative models.</li><li>Safe deployment of models which have the potential to generate harmful content.</li><li>Probing and understanding the limitations and biases of generative models.</li></ul>Excluded uses are described below.<h3 id="heading-26">Out-of-Scope Use</h3>The model was not trained to be factual or true representations of people or events, and therefore using the model to generate such content is out-of-scope for the abilities of this model.<h2 id="heading-27">Limitations and Bias</h2><h3 id="heading-28">Limitations</h3><ul><li>The model does not achieve perfect photorealism</li><li>The model cannot render legible text</li><li>The model struggles with more difficult tasks which involve compositionality, such as rendering an image corresponding to “A red cube on top of a blue sphere”</li><li>Faces and people in general may not be generated properly.</li><li>The autoencoding part of the model is lossy.</li></ul><h3 id="heading-29">Bias</h3>While the capabilities of image generation models are impressive, they can also reinforce or exacerbate social biases.Attribution:"SDXL 0.9 is licensed under the SDXL Research License, Copyright (c) Stability AI Ltd. All Rights Reserved."<h2 id="heading-30">License</h2><a target="_blank" rel="ugc" href="https://huggingface.co/stabilityai/stable-diffusion-xl-base-0.9/blob/main/LICENSE.md">SDXL 0.9 Research License</a>" <a target="_blank" rel="ugc" href="https://huggingface.co/FFusion/FFusionXL-09-SDXL/blob/main/LICENSE.md">FFXL 0.9 Research License</a>"<a target="_blank" rel="ugc" href="mailto:di@ffusion.ai"><img src="https://img.shields.io/badge/Email-di%40ffusion.ai-blue?style=for-the-badge&amp;logo=gmail" alt="Email" /></a><h2 id="heading-31">SAMPLES</h2><a target="_blank" rel="ugc" href="https://cdn-uploads.huggingface.co/production/uploads/6380cf05f496d57325c12194/PxV5UTzx1AYydn9i513ot.png"><img src="https://cdn-uploads.huggingface.co/production/uploads/6380cf05f496d57325c12194/PxV5UTzx1AYydn9i513ot.png" alt="fusion.ai_00093_.png" /></a><a target="_blank" rel="ugc" href="https://cdn-uploads.huggingface.co/production/uploads/6380cf05f496d57325c12194/YlwMyrQvbxXa6KVY2ZSIQ.png"><img src="https://cdn-uploads.huggingface.co/production/uploads/6380cf05f496d57325c12194/YlwMyrQvbxXa6KVY2ZSIQ.png" alt="fusion.ai_00113_.png" /></a><a target="_blank" rel="ugc" href="https://cdn-uploads.huggingface.co/production/uploads/6380cf05f496d57325c12194/ZvtAL425MwNF3mFRukLc-.png"><img src="https://cdn-uploads.huggingface.co/production/uploads/6380cf05f496d57325c12194/ZvtAL425MwNF3mFRukLc-.png" alt="fusion.ai333.png" /></a><a target="_blank" rel="ugc" href="https://cdn-uploads.huggingface.co/production/uploads/6380cf05f496d57325c12194/OMpFvmPKiQwe46Nwbl2f9.png"><img src="https://cdn-uploads.huggingface.co/production/uploads/6380cf05f496d57325c12194/OMpFvmPKiQwe46Nwbl2f9.png" alt="ffusion.aeei.png" /></a>

FFusionXL 0.9 SDXL model + Diffusers

based on SDXL 0.9 Model size3.47B paramsTensor typeF32

Files with SHA256: a05617c461b9e76e1b5998f6c6321ddead56a94a24973ecdf734fe9f3aac0948