<h3 id="heading-69">GAG = Game Asset Generator 📅 Date: <code>2023-10-23T12:19:42</code></h3><h2 id="heading-70">FF_Potion_Generator 🧪⚗️✨</h2><ul><li>Resolution: <code>1024x1024</code></li><li>Architecture: <code>stable-diffusion-xl-v1-base/lora</code></li><li>Network Dim/Rank: <code>32.0</code></li><li>Alpha: <code>1.0</code></li><li>Module: <code>networks.lora</code></li><li>Learning Rates:<ul><li>Overall: <code>0.0005</code></li><li>UNet: <code>0.0005</code></li><li>Text Encoder: <code>5e-05</code></li></ul></li><li>Optimizer: <code>bitsandbytes.optim.adamw.AdamW8bit(weight_decay=0.1)</code></li><li>Scheduler: <code>cosine_with_restarts</code></li><li>Warmup steps: <code>0</code></li><li>Training Details:<ul><li>Epoch: <code>15</code></li><li>Batches per epoch: <code>250</code></li><li>Gradient accumulation steps: <code>1</code></li></ul></li><li>Images:<ul><li>Train images: <code>1000</code></li><li>Regularization images: <code>0</code></li></ul></li><li>Noise &amp; Regularization:<ul><li>Multires noise iterations: <code>6.0</code></li><li>Multires noise discount: <code>0.3</code></li><li>Min SNR gamma: <code>5.0</code></li><li>Zero terminal SNR: <code>True</code></li><li>Max grad norm: <code>1.0</code></li><li>Clip skip: <code>1</code></li></ul></li><li>Dataset: <code>1 Directory</code> with <code>500 images</code></li><li>UNet Weights:<ul><li>Average magnitude: <code>2.9637</code></li><li>Average strength: <code>0.0109</code></li></ul></li><li>Text Encoder Weights:<ul><li>(1) Magnitude: <code>1.7253</code> | Strength: <code>0.0086</code></li><li>(2) Magnitude: <code>1.7251</code> | Strength: <code>0.0067</code></li></ul></li></ul><h3 id="heading-71">🔖 Tags:</h3><pre><code>
potion, painting of one health potion, potion of healing, health potion, magic potions, magical potions, alchemy concept, hyper realistic poison bottle, rendered art, making a potion, potions, fantasy game spell symbol, rpg game item, alchemist bottles, rpg item render, stylized game art, game icon stylized, hyperdetailed scp artifact jar, 3 d icon for mobile game, 3 d render stylized, casting a spell on a potion ... and 904 more!</code></pre> Starting my GAG Serries 🤟 🥃 This one is a Potion LORA for 2.1 SD models. Depending on the model you will get quite different results. <img src="https://imagecache.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/659fb55b-3c4a-49fa-123a-51b622642300/width=525/659fb55b-3c4a-49fa-123a-51b622642300" /><img src="https://imagecache.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/1849160e-202f-4026-e604-043e35305900/width=525/1849160e-202f-4026-e604-043e35305900" /> You can use a background remover while generating and get: <img src="https://imagecache.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/848c54a8-df37-4ff7-8da0-8eb71580d300/width=525/848c54a8-df37-4ff7-8da0-8eb71580d300" /><h3 id="heading-72"> </h3><img src="https://imagecache.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/3e4b1470-9975-4db9-7aff-73f855385600/width=525/3e4b1470-9975-4db9-7aff-73f855385600" /> again this is a test LORA and is heavily dependent on the input model. LyCORIS version will be up for testing in a few hours. And a 2.1 Model is expected in 48-60 hours For PIXALISATION try: <a target="_blank" rel="ugc" href="https://civarchive.com/models/21276/pixhell"> PIXHELL by R6SPY </a>🤟 PS: so far you can use it for demonstration purposes (preview version) with version 1.0 above (lora and models of GAG series) two versions will be available: Licensed - can be licensed for in-game usage upon request (using only my work and licensed images from my end) Community - using sourced free images and sources but you are responsible is used commercially in your project ¯\_(ツ)_/¯ 🤟 🥃

GAG - RPG Potions  |  LoRa XL

Training data: ss_batch_size_per_device: "10",ss_bucket_info: "{"buckets": {"0": {"resolution": [384, 576], "count": 2440}, "1": {"resolution": [512, 512], "count": 49520}, "2": {"resolution": [576, 384], "count": 1720}}, "mean_img_ar_error": 0.0}",ss_bucket_no_upscale: "True",ss_cache_latents: "True",ss_caption_dropout_every_n_epochs: "0",ss_caption_dropout_rate: "0.0",ss_caption_tag_dropout_rate: "0.0",ss_clip_skip: "None",ss_color_aug: "False",ss_enable_bucket: "True",ss_epoch: "2",ss_face_crop_aug_range: "None",ss_flip_aug: "False",ss_full_fp16: "False",ss_gradient_accumulation_steps: "1",ss_gradient_checkpointing: "False",ss_keep_tokens: "0",ss_learning_rate: "0.0001",ss_lowram: "False",ss_lr_scheduler: "linear",ss_lr_warmup_steps: "1074",ss_max_bucket_reso: "1024",ss_max_grad_norm: "1.0",ss_max_token_length: "225",ss_max_train_steps: "10736",ss_min_bucket_reso: "256",ss_mixed_precision: "fp16",ss_network_alpha: "1.0",ss_network_args: "{"conv_dim": "1", "conv_alpha": "1"}",ss_network_dim: "16",ss_network_module: "networks.lora",ss_noise_offset: "0.1",ss_num_batches_per_epoch: "5368",ss_num_epochs: "2",ss_num_reg_images: "0",ss_num_train_images: "53680",ss_optimizer: "lion_pytorch.lion_pytorch.Lion",ss_output_name: "idle.Potionsv_qv",ss_prior_loss_weight: "1.0",ss_random_crop: "False",ss_reg_dataset_dirs: "{}",ss_resolution: "(512, 512)",ss_sd_model_name: "stabilityai/stable-diffusion-2-1",ss_sd_scripts_commit_hash: "00f74d271ab015916a75d6cee6f3af9b0b89f764",ss_seed: "None",ss_session_id: "3030622464",ss_shuffle_caption: "True",ss_text_encoder_lr: "5e-05",ss_total_batch_size: "10",ss_training_comment: "None",ss_training_finished_at: "1679317172.8840992",ss_training_started_at: "1679302821.9685614",ss_unet_lr: "0.0001",ss_v2: "True"

idle_Potion_Generator.yaml

idle_Potion_Generator.safetensors

v 0.1

v0.1 LoRA XL CivitAI ver

USE THE NEW EMBEDDING!!!<h3 id="heading-531">Completed.</h3>Although the model is far from perfect, this is where this particular experiment ends.I'll start over with a different training method.The final version can handle decently 1024x1024 resolution for most of its concepts.Right now I don't feel like posting the "keywords" the model was trained on, maybe I'll suffice later.

AllWorkForkRowk_InProgress_(SD2.1)

Overall improvement, overfitting in the characters it had learned about, I'll fix it later and provide the list of the characters.

allworkforkrowk_v02GammaFasterlr512.yaml

allworkforkrowk_v02GammaFasterlr512.safetensors

V0.2_Gamma_(FasterLR)_512

v1.0_768x1024 Definitive

V0.1_Beta_512

V0_Alpha_tunedAt512

Trained 5 epochs on Illuminati Diffusion v1.1 Epoch 1 and 4 are shit Rest of them are really good, epoch 5 i recommend using at a lora strength of 0.7 and below because i just had one image with the car next to a fence a house and some trees and most of the time it's trying to replicate the surroundings in that image which is weird and also makes the car look hideous, maybe it's overcooked. &lt;0.7 fixes this issue, the car looks great still. Images are generated on Illuminati Diffusion v1.1 ps: I will try to add the rest of the good epochs later, civitai is acting up when i upload images...

Speedtail

Speedtail-000002.safetensors

Speedtail-000002.yaml

epoch 3

epoch 2

epoch 5

Trained 5 epochs on Illuminati Diffusion v1.11st is shit2, 3, 4 are ok5th is the best i thinkcan also use front view, top view, rear view, side viewclose-up(rear view) and distant shot(front view) might also work

McLaren F1 LM

F1 LM-000004.yaml

F1 LM-000004.safetensors

Decided to make a lora of a car, trained on Illuminati Diffusion v1.1

Amg One

Amg One.yaml

Amg One.safetensors

About model A simple model trained on a custom dataset containing over 100 coloring book type images.If you enjoy this model and would like me to improve on it, <a target="_blank" rel="ugc" href="https://www.buymeacoffee.com/mrhup">buy me a coffe </a>☕ Installation:Download both the ckpt and yaml files. Ensure that the same naming pattern is used and copy them under models/Stable-Diffusion path in your local/cloud SD installation. Stable Diffusion 2.1 is required for the model to work correctly. Black images issue:2.1 models need to have a web-ui config modified - if you are getting black images - go to your config file and add to COMMANDLINE_ARGS= <code>--no-half</code> - potentially it could work with <code>--xformers</code> instead (if supported). This line might slow your generations a bit but will not affect negatively your output.Prompt suggestion:<code>bichon havanese wearing sunglasses COLR_001, (((white background))), coloring book, line art, high resolution, black and white, colorless</code>Negative: <code>((watermark)), (text), color, shading, gradient, shadows, transparency, noisy, blurred</code>

coloring-book

coloringBook_coloringBook.ckpt

coloringBook_coloringBook.yaml

This was my first attempt at creating a separate Lycoris(LoHa) using the dataset from my CharHelper model + an extra hundred images. The biggest challenge was not having the mouth and teeth look like a janky mess. This model was the setup that gave me the best results. It is a LoHa trained with a 128 network rank 5 steps each for 22 epochs WITH 183 regularization images. The regularization images were important for achieving the closest likenesses. Note: the model was not trained with any nudity, however, due to most orcs in the dataset being topless, a female Orc may show some nudity. To bypass this you can put "shirtless" or "topless" in the negative prompt. Some useful keywords to activate the model are "CHV3COrc, Orc, Orcish, and Orc theme"

Orcish

Orc_reg_128.yaml

Orc_reg_128.safetensors

Orcish V1.0

My firs LoRa, apologize if it's not perfect. All training suggestions are welcome. Trained on 110 pictures, base modelsrecommendations for SD 1.5 EULER A higher step number (25-50) CFG 6-7 768x512recommendations for SD 2.1 768x (don't forget to download .yaml) EULER A low step number (around 20-30) CFG 6 or lower 960x576 or 1024x768 decrease LoRa weight if needed

Citroen C6

CitroenC6-21-v1.yaml

CitroenC6-21-v1.safetensors

SD 2.1 768x v1.0

SD 1.5 v1.0

📣 FFUSION AI SD2.1- 768 BaSE Public 1.0.0Release is Here! Diffusers available at <a target="_blank" rel="ugc" href="https://huggingface.co/FFusion">https://huggingface.co/FFusion</a><img src="https://cdn-uploads.huggingface.co/production/uploads/6380cf05f496d57325c12194/FHidip3rT8mL1UdGDyfSd.jpeg" alt="ffusion-basesm32.jpg" /><h1 id="heading-1523">STABLE DIFFUSION 2.1 768+ MODEL</h1>before complaining about usage if you haven't used 2.1 stick to 1.5 models<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/200d110a-afe6-47d9-b3d0-9e1f66f3cd12/width=525/200d110a-afe6-47d9-b3d0-9e1f66f3cd12.jpeg" /><img src="https://static.xx.fbcdn.net/images/emoji.php/v9/t1f/1.5/16/1f680.png" alt="🚀" /> Introducing <a target="_blank" rel="ugc" href="http://FFusion.AI">FFusion.AI</a>-beta-Playground on Hugging Face Spaces!<a target="_blank" rel="ugc" href="https://huggingface.co/spaces/FFusion/FFusion.AI-beta-Playground">https://huggingface.co/spaces/FFusion/FFusion.AI-beta-Playground</a> <a target="_blank" rel="ugc" href="https://ffusion.ai/">https://ffusion.ai/</a>──────────────────────────────────────We're thrilled to announce the launch of our new application, <a target="_blank" rel="ugc" href="http://FFusion.AI">FFusion.AI</a>-beta-Playground, now live on Hugging Face Spaces! This cutting-edge tool harnesses the power of AI to generate stunning images based on your prompts. <img src="https://static.xx.fbcdn.net/images/emoji.php/v9/tdb/1.5/16/1f3a8.png" alt="🎨" /><img src="https://static.xx.fbcdn.net/images/emoji.php/v9/ta7/1.5/16/1f5bc.png" alt="🖼️" />──────────────────────────────────────With <a target="_blank" rel="ugc" href="http://FFusion.AI">FFusion.AI</a>-beta-Playground, you can:<img src="https://static.xx.fbcdn.net/images/emoji.php/v9/t93/1.5/16/31_20e3.png" alt="1️⃣" /> Generate images from a variety of pre-trained models including <a target="_blank" rel="ugc" href="http://FFUSION.ai">FFUSION.ai</a>-768-BaSE, <a target="_blank" rel="ugc" href="http://FFUSION.ai">FFUSION.ai</a>-v2.1-768-BaSE-alpha-preview, and <a target="_blank" rel="ugc" href="http://FFusion.ai">FFusion.ai</a>.Beta-512.<img src="https://static.xx.fbcdn.net/images/emoji.php/v9/tb2/1.5/16/32_20e3.png" alt="2️⃣" /> Experiment with different schedulers to fine-tune the image generation process.<img src="https://static.xx.fbcdn.net/images/emoji.php/v9/td1/1.5/16/33_20e3.png" alt="3️⃣" /> View the generated images right in your browser and save them for later use.──────────────────────────────────────Our application is built on top of the diffusers library and uses StableDiffusionPipeline for image generation. It's powered by Gradio for a user-friendly interface. And here's the exciting part: very soon, it will run on a CUDA-enabled environment for optimal performance, thanks to our partners at RUNPOD! <img src="https://static.xx.fbcdn.net/images/emoji.php/v9/te5/1.5/16/1f4bb.png" alt="💻" /><img src="https://static.xx.fbcdn.net/images/emoji.php/v9/t1f/1.5/16/1f680.png" alt="🚀" />Stay tuned for this upcoming enhancement that will take your image generation experience to the next level. We're thrilled to be partnering with <a target="_blank" rel="ugc" href="http://RUNPOD.io">RUNPOD.io</a> to bring you this cutting-edge technology.──────────────────────────────────────To get started, simply enter your prompt, select the models you want to use, choose a scheduler, and let our application do the rest.──────────────────────────────────────Check out <a target="_blank" rel="ugc" href="http://FFusion.AI">FFusion.AI</a>-beta-Playground now at FFusion/<a target="_blank" rel="ugc" href="http://FFusion.AI">FFusion.AI</a>-beta-Playground and start creating your own unique images today! <img src="https://static.xx.fbcdn.net/images/emoji.php/v9/te5/1.5/16/1f389.png" alt="🎉" /><img src="https://static.xx.fbcdn.net/images/emoji.php/v9/te5/1.5/16/1f389.png" alt="🎉" />──────────────────────────────────────We're excited to see what you'll create with <a target="_blank" rel="ugc" href="http://FFusion.AI">FFusion.AI</a>-beta-Playground. Your feedback is invaluable to us, so please don't hesitate to share your thoughts and suggestions. Enjoy exploring the possibilities of AI-powered image generation! <img src="https://static.xx.fbcdn.net/images/emoji.php/v9/t95/1.5/16/1f4a1.png" alt="💡" /><img src="https://static.xx.fbcdn.net/images/emoji.php/v9/t39/1.5/16/1f31f.png" alt="🌟" />🔭 We are thrilled to launch the public beta release of FFUSION Ai, though we want to clarify that it's currently limited in its breadth. Having been trained on just a fraction of our full image collection (20%), the capabilities of the model are not yet fully realized. This early version is primarily intended for experimentation with various prompt combinations and initial testing.💡 While we're committed to delivering the highest level of excellence, we want to highlight that our model, notably the Unet component, is still developing its proficiency with certain objects and faces. But fear not, we're actively fine-tuning these areas as we progress towards the final release.🙏 A huge shout out to our Reddit community for their support in alpha testing and for helping the text encoder respond to some exciting fuse ideas. We couldn't have come this far without you!💡 Your contribution in this beta testing phase is extremely crucial to us. We invite you to explore the model extensively, experiment with it, and do not hesitate to report any prompts that don't meet your expectations. Your feedback is our guiding light in refining the performance and overall quality of FFUSION Ai.⚠️ Attention: The model is based on Stable Diffusion 2.1 - 512 and is designed for optimal performance up to a resolution of approximately 600-700 pixels. For larger image sizes, we recommend upscaling them independently or patiently waiting for our final release that's just around the corner. This forthcoming release will enhance performance and support for higher resolutions.👥 Thank you for being part of the FFUSION Ai beta testing community. Your support, feedback, and passion inspire us to continually develop a pioneering tool that is set to revolutionize creativity and visualization. Together, we can shape the future of storytelling and creativity.🔮 Why not add some effects to your favorite prompts or fuse them together for a surreal twist? (Please note, Pen Pineapple Apple Pan effects and FUSIONS are excluded in this beta version)🔒 With over 730.9449 hours of dedicated training sessions, our Fusion AI model offers a wealth of data subsets and robust datasets developed in collaboration with two enterprise corporate accounts for Mid Journey. We also pride ourselves in having an effective utilization of GPU usage, making the most out of our partnership with Idle Stoev, Source Code Bulgaria, Praesidium CX &amp; BlackSwan Technologies. 🚀Full transparency on our extensive 700,000-image dataset, training methodologies, classifications, and successful experiments is on its way. This information will be released shortly after the final version, further establishing FFUSION Ai as a trusted tool in the world of AI-powered creativity. Let's continue to imagine, create and explore together! <img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/b66ba685-f8ed-4e9b-be34-0aece1ebd993/width=525/b66ba685-f8ed-4e9b-be34-0aece1ebd993.jpeg" /> Model Overview: Unleashing the Power of Imagination!FFUSION AI is a state-of-the-art image generation and transformation tool, developed around the leading Latent Diffusion Model. Leveraging Stable Diffusion 2.1, FFUSION AI converts your prompts into captivating artworks. Discover an imaginative landscape where ideas come to life in vibrant, surreal visuals.<ul><li>Developed by: Idle Stoev, Source Code Bulgaria, Praesidium CX &amp; BlackSwan Technologies</li><li>Shared by: FFusion AI</li><li>Model type: Diffusion-based text-to-image generation model</li><li>Language(s) (NLP): English</li><li>License: CreativeML Open RAIL++-M License</li></ul><h2 id="heading-1524">Model Use: Enabling Creativity and Exploring AI Frontiers</h2>Designed for research and artistic exploration, FFUSION AI serves as a versatile tool in a variety of scenarios:<ul><li>Investigating biases and constraints in generative models</li><li>Unleashing creativity in artistic endeavors</li><li>Infusing AI-driven innovation into educational or creative tools</li><li>Furthering research in the exciting field of generative models</li><li>Repository: <a target="_blank" rel="ugc" href="https://github.com/1e-2">https://github.com/1e-2</a></li><li>Demo: <a target="_blank" rel="ugc" href="https://huggingface.co/spaces/FFusion/FFusion.AI-beta-Playground">https://huggingface.co/spaces/FFusion/FFusion.AI-beta-Playground</a></li></ul>Out-of-Scope Use and Prohibited Misuse:<ul><li>Generating factually inaccurate representations of people or events</li><li>Inflicting harm or spreading malicious content such as demeaning, dehumanizing, or offensive imagery</li><li>Creating harmful stereotypes or spreading discrimination</li><li>Impersonating individuals without their consent</li><li>Disseminating non-consensual explicit content or misinformation</li><li>Violating copyrights or usage terms of licensed material</li></ul><h2 id="heading-1525">Model Limitations and Bias</h2>While our model brings us closer to the future of AI-driven creativity, there are several limitations:<ul><li>Achieving perfect photorealism or surrealism is still an ongoing challenge.</li><li>Rendering legible text could be difficult without further ~30min training on your brand.</li><li>Accurate generation of human faces, especially far away faces, is not guaranteed (yet).</li></ul><h2 id="heading-1526">Model Releases</h2>We are thrilled to announce:<ul><li>Version 512 Beta: Featuring LiTE and MiD BFG model variations</li><li>Version 768 Alpha: BaSE, FUSION, and FFUSION models with enhanced training capabilities, including LoRa, LyCORIS, Dylora &amp; Kohya-ss/sd-scripts.</li><li>Version 768 BaSE: A BaSE Ready model for easy applying more than 200 build op LoRA models trained along the way.</li></ul><h2 id="heading-1527">Environmental Impact</h2>In line with our commitment to sustainability, FFUSION AI has been designed with carbon efficiency in mind:<ul><li>Hardware Type: A100 PCIe 40GB</li><li>Hours used: 1190</li><li>Cloud Provider: CoreWeave &amp; Runpod (official partner)</li><li>Compute Region: US Cyxtera Chicago Data Center - ORD1 / EU - CZ &amp; EU - RO</li><li>Carbon Emitted: 124.95 kg of CO2 (calculated via Machine Learning Impact calculator)</li></ul>That said all LoRA and further models are based on initial training.<h2 id="heading-1528">Model Card Authors</h2>This model card was authored by Idle Stoev and is based on the Stability AI - Stable Diffusion 2.1 model card.<h2 id="heading-1529">Model Card Contact</h2><a target="_blank" rel="ugc" href="mailto:di@ffusion.ai">di@ffusion.ai</a>Download the <a target="_blank" rel="ugc" href="https://huggingface.co/FFusion/FFusion-BaSE/blob/main/FFusion-BaSE.safetensors">FFUSION AI diffusers - 768 BaSE Release here</a>. <h1 id="heading-81">🔬 Intended Use: From Research to Artistry 🎨</h1><img src="https://s3.amazonaws.com/moonup/production/uploads/6380cf05f496d57325c12194/PKYptThtja5IkYSL9ERTt.jpeg" alt="Screenshot_1002.jpg" /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/c34e6764-d4ca-4b2c-8731-9ba8d8f056fd/width=525/c34e6764-d4ca-4b2c-8731-9ba8d8f056fd.jpeg" />

FFUSION AI - SD 2.1

Model Overview: Unleashing the Power of Imagination!FFUSION AI is a state-of-the-art image generation and transformation tool, developed around the leading Latent Diffusion Model. Leveraging Stable Diffusion 2.1, FFUSION AI converts your prompts into captivating artworks. Discover an imaginative landscape where ideas come to life in vibrant, surreal visuals.<ul><li>Developed by: Idle Stoev, Source Code Bulgaria, Praesidium CX &amp; BlackSwan Technologies</li><li>Shared by: FFusion AI</li><li>Model type: Diffusion-based text-to-image generation model</li><li>Language(s) (NLP): English</li><li>License: CreativeML Open RAIL++-M License</li></ul>Model Use: Enabling Creativity and Exploring AI FrontiersDesigned for research and artistic exploration, FFUSION AI serves as a versatile tool in a variety of scenarios:<ul><li>Investigating biases and constraints in generative models</li><li>Unleashing creativity in artistic endeavors</li><li>Infusing AI-driven innovation into educational or creative tools</li><li>Furthering research in the exciting field of generative models</li><li>Repository: <a target="_blank" rel="ugc" href="https://github.com/1e-2">https://github.com/1e-2</a></li><li>Demo: <a target="_blank" rel="ugc" href="https://huggingface.co/spaces/FFusion/FFusion.AI-beta-Playground">https://huggingface.co/spaces/FFusion/FFusion.AI-beta-Playground</a></li></ul>Out-of-Scope Use and Prohibited Misuse:<ul><li>Generating factually inaccurate representations of people or events</li><li>Inflicting harm or spreading malicious content such as demeaning, dehumanizing, or offensive imagery</li><li>Creating harmful stereotypes or spreading discrimination</li><li>Impersonating individuals without their consent</li><li>Disseminating non-consensual explicit content or misinformation</li><li>Violating copyrights or usage terms of licensed material</li></ul>Model Limitations and BiasWhile our model brings us closer to the future of AI-driven creativity, there are several limitations:<ul><li>Achieving perfect photorealism or surrealism is still an ongoing challenge.</li><li>Rendering legible text could be difficult without further ~30min training on your brand.</li><li>Accurate generation of human faces, especially far away faces, is not guaranteed (yet).</li></ul>Model ReleasesWe are thrilled to announce:<ul><li>Version 512 Beta: Featuring LiTE and MiD BFG model variations</li><li>Version 768 Alpha: BaSE, FUSION, and FFUSION models with enhanced training capabilities, including LoRa, LyCORIS, Dylora &amp; Kohya-ss/sd-scripts.</li><li>Version 768 BaSE: A BaSE Ready model for easy applying more than 200 build op LoRA models trained along the way.</li></ul>Environmental ImpactIn line with our commitment to sustainability, FFUSION AI has been designed with carbon efficiency in mind:<ul><li>Hardware Type: A100 PCIe 40GB</li><li>Hours used: 1190</li><li>Cloud Provider: CoreWeave &amp; Runpod (official partner)</li><li>Compute Region: US Cyxtera Chicago Data Center - ORD1 / EU - CZ &amp; EU - RO</li><li>Carbon Emitted: 124.95 kg of CO2 (calculated via Machine Learning Impact calculator)</li></ul>That said all LoRA and further models are based on initial training.Model Card AuthorsThis model card was authored by Idle Stoev and is based on the Stability AI - Stable Diffusion 2.1 model card.Model Card Contact<a target="_blank" rel="ugc" href="mailto:di@ffusion.ai">di@ffusion.ai</a>Download the <a target="_blank" rel="ugc" href="https://huggingface.co/FFusion/FFusion-BaSE/blob/main/FFusion-BaSE.safetensors">FFUSION AI - 768 BaSE diffusers @ Huggingface</a>

ffusionAISD21_baseV100.ckpt

ffusionAISD21_baseV100.safetensors

ffusionAISD21_baseV100.yaml

BaSE - v1.0.0

768-BaSE-Alpha

512-beta-BFG-build.0401

MiD-512-beta-build0401

LiTE-512-beta-build0201

The beggining of a great project !Note : don't generate BELOW the base resolution (768*768)<h2>Providence - Prelude  </h2><ul><li>A precise chosen noise offset setting</li><li>A large variety of subjects, I took care to to keep the flavour at maximum from genesis seeds feel</li><li>No baked VAE (works fine without any tbh)</li></ul><h3>Hyper Realism</h3><img src="https://imagecache.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/a306da42-77a0-4827-ffc3-c65498c5a200/width=525/a306da42-77a0-4827-ffc3-c65498c5a200" />For good results please use the <a target="_blank" rel="ugc" href="https://civarchive.com/models/15432/forge-negatives-and-helpers-21">forAnalogPortrait</a> negative embedding.<h3>Digital painting</h3><img src="https://imagecache.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/c040c883-4ecb-4a57-5326-8fe3c130cf00/width=525/c040c883-4ecb-4a57-5326-8fe3c130cf00" />For my digital painting finetune, please use “character portrait” along “digital painting” for enhance the quality<h3>Analog</h3><img src="https://imagecache.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/3f67c9cf-f618-4792-3c30-da77b0fc0f00/width=525/3f67c9cf-f618-4792-3c30-da77b0fc0f00" />use the <a target="_blank" rel="ugc" href="https://civarchive.com/models/15432/forge-negatives-and-helpers-21">forAnalogPortrait</a> negative embedding aswell.<h2>Troubleshoot :</h2>NansException: A tensor with all NaNs was produced in Unet. This could be either because there's not enough precision to represent the picture, or because your video card does not support half type. Try setting the "Upcast cross attention layer to float32" option in Settings &gt; Stable Diffusion or using the --no-half commandline argument to fix this. Use --disable-nan-check commandline argument to disable this check.A 2.1 model require either --xformers and / or --no-half start argument from the webui-user.batset COMMANDLINE_ARGS= --no-half --xformerscheck also if your config file is named correctly alongside the model<img src="https://imagecache.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/dc044796-0893-4d54-24bb-715c30dd9500/width=525/dc044796-0893-4d54-24bb-715c30dd9500.jpeg" />Happy tuning ! <a target="_blank" rel="ugc" href="http://romerorz.art">romerorz.art</a> - <a target="_blank" rel="ugc" href="https://twitter.com/romero_erzede">Twitter </a>- <a target="_blank" rel="ugc" href="https://www.deviantart.com/romerorz">DeviantArt</a> - <a target="_blank" rel="ugc" href="https://romerorz.gumroad.com/">Gumroad</a>

Providence

providence_2112Anchor.yaml

providence_2112Anchor.safetensors

2.1 - 1.2 - Anchor

2.1 - 1.1 - Anchor

2.1 - 1.0 - Prelude

Made a lora of the lamborghini countach lp800, trained on Illuminati Diffusion v1.1don't know why but it's more prone to showing the car from the back, just keep trying until you get it from the front, i made 3 epochs and this is the first one, the second one was a bit unrealistic, and the third one was stylizing the car a bit too much, especially on the rear end, this epoch sticks with the normal layout of the taillights compared to the third epoch.images are generated on Illuminati Diffusion v1.1

Countach LP800

Countach lp800.safetensors

Countach lp800.yaml

Speedtail-000004.safetensors

Speedtail-000004.yaml

📣 FFUSION AI - 768 BaSE Public alpha Release is Here!download the <code>di.FFUSION.ai-v2.1-768-BaSE-alpha </code> 🧨<code>DIFFUSERS</code> at <a target="_blank" rel="ugc" href="https://huggingface.co/FFusion/">https://huggingface.co/FFusion/</a>

ffusionAISD21_768BaseAlpha.ckpt

ffusionAISD21_768BaseAlpha.safetensors

ffusionAISD21_768BaseAlpha.yaml

Tilt Shift LoHa is a LyCORIS/LoHa trained with v2.1 as a base. It can produce some pretty amazing aerial tilt shift photography on all different types of landscapes. It is not the greatest at toon-like JRPG landscapes as it's main dataset is real photographs. Edit: For a more cartoony style, use with <a target="_blank" rel="ugc" href="https://civarchive.com/models/10003?modelVersionId=79886">https://civarchive.com/models/10003?modelVersionId=79886</a>

Tilt Shifted LoHa

Tilt Shift LoHa is a LyCORIS/LoHa trained with v2.1 as a base. It can produce some pretty amazing aerial tilt shift photography on all different types of landscapes. It is not the greatest at toon-like JRPG landscapes as it's main dataset is real photographs. Edit: For a more cartoony style, use with <a target="_blank" rel="ugc" href="https://civitai.com/models/10003?modelVersionId=79886">https://civitai.com/models/10003?modelVersionId=79886</a>

Tilt Shift LoHa.safetensors

Tilt Shift LoHa.yaml

v1.0

About this versionDont forget to download the config fileRecommended render resolutions: 960 x 576 or 1024 x 768 Basic settings should work fine, like EULER A, CFG 6-7 Steps 20-30 (...100) For more complet prompts you might need to add extra weight for the trigger word, or increase TE weight for the LoRa. All feedback are welcome

Citroen C5 (x7)

Dont forget to download the config fileRecommended render resolutions: 960 x 576 or 1024 x 768 Basic settings should work fine, like EULER A, CFG 6-7 Steps 20-30 (...100) For more complet prompts you might need to add extra weight for the trigger word, or increase TE weight for the LoRa. All feedback are welcome

c5x7-21-v01.yaml

c5x7-21-v01.safetensors

SD 2.1 768 v1.0

Anyone that has ever tried to generate an image of ATJ has found no joy in the resulting image. I'm here to fix all that. just plug in the LoRa and craft your prompts. It is by no means perfect but it looks more like her than the stock version embedded in the SD 1.x and 2.x checkpoints. Please be mature about your generations with my models folks, I don't want to see any NSFW plastered everywhere of ATJ. If you want to generate NSFW images with ATJ, just keep them to yourself, OK? I'm sure you wouldn't want fake nudes of your sister or yourself plastered over the internet, would you?

Anya Taylor-Joy 2.x

AnyTayJoy-woman_Vparam_Reged.yaml

AnyTayJoy-woman_Vparam_Reged.safetensors

First of a series of EVO brand cars LoRA models.All training suggestions are welcome.v0.741 first SDXL version, includes EVO 3 car, Cattedrale di Palermo square, Teatro Massimo &amp; Palermo streets

EVO 3

Dataset: 44 images (x 30 repeats)BLIP captioning1320 cherry picked reg images D-AdaptationLR: 0.5Constant with warmup (10%)Train batch size: 4

 evo_car_D-adaptation-LoRA_SD21-768_v044.yaml

 evo_car_D-adaptation-LoRA_SD21-768_v044.safetensors

v0.44 (SD2.1 768)

v0.85 Milazzo (SDXL)

v0.85 Ragusa (SDXL)

v0.83 Agrigento (SDXL)

v0.741 Palermo (SDXL)

v0.61 Poli+Lecce (SD1.5)

v0.55 Polignano (SD1.5)

v0.45 Lecce (SD2.1 768)

v0.3 (SD2.1 768)

v0.2 (SD 2.1 768)

Multi concepts LoRA: evo car + dm plaza (piazza Duomo Lecce concept)Dataset evo: 44 images (x 30 repeats), dm: 29 images (x30 repeats)BLIP captioning1320 car (cherry picked) reg images mix, 232 plaza reg imagesD-AdaptationLR: 0.5

 evo_car-dm_plaza_D-adaptation-LoRA_SD21-768_v045.yaml

 evo_car-dm_plaza_D-adaptation-LoRA_SD21-768_v045.safetensors

Heavy face and standart pose variety fixA lot less overfittingGreat as “raw” without heavy negatives

providence_2111Anchor.safetensors

providence_2111Anchor.yaml

Trained on a dataset of 118 hi-res images (wiith BLIP captions manually edited) with Repeats = 20 and 2360 regularisation images, source model: stabilityai/stable-diffusion-2-1 with v2 and v_parameterization enabled.Relevant Kohya settings:<ul><li>Learning rate: 0.0001</li><li>LR Scheduler: cosine</li><li>LR warmup (% of steps): 10</li><li>Optimizer: AdamW8bit</li><li>Text Encoder learning rate: 0.00005</li><li>Unet learning rate: 0.0001</li><li>Network Rank (Dimension): 128</li><li>Network Alpha: 1</li><li>Max resolution: 768,768</li><li>Clip skip: 1</li><li>Min SNR gamma: 0 </li></ul>

evo-car-lora-sd21-768_v02.yaml

evo-car-lora-sd21-768_v02.safetensors

This is a Locon. You probably need the extension in A1111.Config file is in the "Text Encoder" download.The base model I used is <a target="_blank" rel="ugc" href="https://civarchive.com/models/36152/illuminutty-diffusion">Illuminutty Diffusion</a>You need to download the negative embeddings to replicate my generations. <a target="_blank" rel="ugc" href="https://files.catbox.moe/0gjduq.zip">https://files.catbox.moe/0gjduq.zip</a>You may want to look at the tags in the dataset I provided to figure out what prompts to use.Trained on images of stained glass from League of Legends' <a target="_blank" rel="ugc" href="https://www.youtube.com/watch?v=JKnM6Hh5gOg&amp;t=0s&amp;ab_channel=LeagueofLegends">Shurima Cinematic</a>Dataset is poorly tagged, if someone can help me manually tag the dataset that would be awesome. It would help me train a better model.

Shurima Cinematic - Stained Glass

This is a Locon. You probably need the extension in A1111.

shurima-1-000010.yaml

shurima-1.safetensors

image.zip

providence_2110Prelude.yaml

providence_2110Prelude.safetensors

It's my original concept/style, aspects of gold, godhood and cosmic terror - gods in gold. Using it at strength 1 is fine, no overcook here. It's trained on 768x768 cap, so generations are best at this base resolution.

RFKTR's gods in gold [SD2.1 x768]

gdsingld.safetensors

gdsingld.yaml

Trained 3 epochs in Illuminati Diffusion v1.1 3rd at lora strength 1 is the only one that can produce stylized windshield background without sticking with the default trained ones. 1 and 2 suffer from that :( You can get some really good results even at 0.6 if i remember correctly 3rd is the best one in my opinion.If you only use "amdbx interior" you might get some funky results, for example i tried having a (city view:1.5) outside my windshield without inputting "car interior" in the prompt and generated some really weird images and a basic kitchen with a window view inside a city

Aston Martin DBX Interior

Aston Martin DBX Interior-000002.safetensors

Aston Martin DBX Interior-000002.yaml

Trained 15 images on Illuminati Diffusion v1.1 trained 3 epochs, all 3 are ok, second and third are i think better, second being the best, both first and second make great wheels too, lower the lora if you want more artistic freedom Images generated on Illuminati Diffusion v1.1

Ferrari 360 Modena Liberty Walk

LBWK360MODENA-000001.yaml

LBWK360MODENA-000001.safetensors

Training notes:V0.6 (SDXL) same "cooking" settings as the Ferrari 312 B v0.6 LoRAV0.4 (SDXL) was trained on 38 imgs (x 4 repeats) with manually reviewed WD14 captions and 152 real 1970s formula one season photos as regularisation images, against sdXL_v10VAEFix model.V0.2 (SDXL) FAT Bigger 1.7 GB version but generate more realistic imagesAnyway for this vehicle I really need a better dataset with more hi-res picsThe McLaren M14A is a Formula One racing car built and raced by McLaren in the 1970 World Championship and the 1971 World Championship.The M14A was an evolution of the previous M7A and M7C, with the primary change being the rear brakes were mounted inboard instead of outboard. As with the M7, the M14A was powered by a Cosworth DFV V8 and a Hewland 5-speed manual gearbox.The unfortunate 1970s season began with two second places, one fourth and three retirements for Bruce McLaren and Denny Hulme.

McLaren M14A Formula One (1970)

This is the result of my 5th training run for the McLaren M14A. I used 64 images for 64 steps and 4096 classification images.Recommended LoRA weight : 0.8 - 1.0Trigger words: m14a f1 car 70sThis LoRA model has been trained to generate images on Stable Diffusion 2.1 768

m14af1car70s.safetensors

m14af1car70s.yaml

v0.5 (SD 2.1 768)

v0.6 (SDXL) LAST VERSION

v0.4 (SDXL)

v0.2 (SDXL) FAT 1.7 GB

v0.3 (SD 2.1 768)

First of a series of DR brand cars LoRA models. All training suggestions are welcome.Trained on a 38 images dataset and with 3800 reg images, model SD 2.1 768

DR DR1.0 EV (city car)

Trained on a 38 images dataset and with 3800 reg images, model SD 2.1 768

drccitycar_v0.3.yaml

drccitycar_v0.3.safetensors

v0.3 epoch 3 (SD 2.1 768)

v0.1(SD 2.1 768)

Trained 3 epochs on Illuminati Diffusion v1.11st is bad. 2nd epoch is really artistic and sometimes adds tuning to the car at lower weight and looks really cool, has trouble with pop-up headlights. 3rd epoch is very realistic, only tested at weight 1, looks good.Images generated on Illuminati Diffusion v1.1

RX 7 FD

Rx7FD-000002.safetensors

Rx7FD-000002.yaml

Speedtail-000001.yaml

Speedtail-000001.safetensors

Trained 3 epochs on Illuminati Diffusion v1.1 1st is shit2nd and 3rd epoch are good, indistinguishable i'd say.... ok maybe the 3rd epoch has a more sharp rear end. lower the lora strength for more artistic vibe.

Aston Martin DBX

Aston Martin DBX-000002.safetensors

Aston Martin DBX-000002.yaml

<ul><li>现在来到了200w steps loss曲线仍在降低虽然很慢,同时我的偏置噪声似乎设定的有些高了.初始Noise offset 0.05.从画面图像来看有部分颜色有些失真或溢出..比如绿色（。也许有人能用lora合并来修复画面问题。(在sd2.1的fintune模型上lora好像泛用性较低,也许需要在该模型上训练的才会有效果）我尝试了一个sd2.1的lora发现对画面没有起效，不知道原因在哪</li><li>Now we have reached 200w steps. The curve loss is still decreasing, though very slowly. Meanwhile, my bias noise seems to have been set too high. Initial Noise offset 0.05. Judging from the image, some colors are distorted or overflow.. For example, green (. Maybe someone can fix the graphics with lora merge.(lora seems to be less general on sd2.1 fintune model, maybe it needs to be trained on that model to be effective) I tried an sd2.1 lora and found it didn't work on the screen, I don't know why</li><li>这个模型在3.1w张niji图像训练了200个gpu小时合计120w steps。但是loss任未收敛至合适范围.同时由于训练文本倾向原因导致模型短语理解能力失调.可能需要重新调整.也就是说这个模型不能通过简单的1girl出图.有兴趣的人可以在上面继续做微调，亦或是能给我提供补充数据集也欢迎联系我</li><li>This model was trained on 3.1w niji images for 200 gpu hours totaling 120k steps. However, the loss does not converge to the appropriate range. At the same time, the training text tendency causes the dyscomprehension of model phrases. May need to readjust. That is to say, this model cannot be mapped by a simple 1girl. If you are interested, please feel free to contact me for further fine-tuning, or if you can provide me with supplementary data sets</li></ul>

anime2.1

现在来到了200w steps loss曲线仍在降低虽然很慢,同时我的偏置噪声似乎设定的有些高了.初始Noise offset 0.05.从画面图像来看有部分颜色有些失真或溢出..比如绿色（。也许有人能用lora合并来修复画面问题。(在sd2.1的fintune模型上lora好像泛用性较低,也许需要在该模型上训练的才会有效果）我尝试了一个sd2.1的lora发现对画面没有起效，不知道原因在哪

anime21_v15.safetensors

anime21_v15.yaml

v1.5

Aston Martin DBX-000001.safetensors

Aston Martin DBX-000001.yaml

Rx7FD-000001.safetensors

Rx7FD-000001.yaml

All models for pixelart I've tried did not give me what I wanted so I went and made my own. The model recreates Pixelart style but with more modern take on the 8bit style. This model is great for side scroller environments but it works great for characters when using controlnet for 2.1 for openpose and also it's great for props and other things.Enjoy!Official YouTube Channel:<a target="_blank" rel="ugc" href="https://www.youtube.com/@spybgtoolkit"> [CLICK HERE]</a>Patreon: <a target="_blank" rel="ugc" href="https://www.patreon.com/SPYBGToolkit">[CLICK HERE]</a>

PIXHELL

<ul><li>added config file (for everyone who were asking me)</li><li>added fp16 version of safetensor model so now it should work for people with slower machines with less vram</li></ul>

pixhell_v20.safetensors

pixhell_v20.yaml

v 2.0

v 1.0

I also publish this previous version because despite being less accurate it manages to generate unusual images

mcaf1car70s.safetensors

mcaf1car70s.yaml

I was attempting to make a lora that would see the horror side of prompts. This is the result, It is not exactly what I was going for, but I like the images it creates. It is best to add ",by daveScare" to the end of a prompt if your not prompting for anyhting sinister... it should pick up on sinister prompts itself. Been using this in comfyui, be sure to play around with the clip and model strengths to get your desired effect EDIT: <s>its a 2.1 Lora so may need the yaml file.....How ever Civitai does not allow me to upload one. </s> Yaml now included in the downloads

Files with SHA256: 0e3db941233ede7ddfb365f5f6d29fb05f767eb8e5e602cd83bb17d905da1f8e