<h2 id="transform-your-images-with-natural-language-instructions-isa08z788"><span style="color:rgb(117, 28, 28)">🔮 </span><span style="color:rgb(0, 178, 255)">Transform Your Images with Natural Language Instructions</span></h2><p>Used with FluxFill-Dev The perfect companion for seamless editing</p><p><a target="_blank" rel="ugc" href="https://civarchive.com/models/1534861/icedit-workflow"><strong><span style="color:rgb(255, 0, 242)">📌 </span><span style="color:rgb(0, 223, 252)">Complete Workflow</span></strong></a></p><p></p><h2 id="overview:-the-future-of-image-editing-is-here-qma5holu8"><span style="color:rgb(255, 0, 191)">📝 Overview: The Future of Image Editing is Here</span></h2><p>ICEdit-MoE-LoRA represents a breakthrough in AI image editing - a lightweight Low-Rank Adaptation (LoRA) module that brings professional-grade, instruction-based editing capabilities to large diffusion transformers. By intelligently fine-tuning only 1% of parameters using just 0.5% of the original training data, it achieves extraordinary editing capabilities that match or even surpass premium commercial solutions.</p><h2 id="-dnlzo6hw9"></h2><h2 id="complete-installation-guide-zigfkxzon"><span style="color:rgb(255, 0, 225)">🛠️ Complete Installation Guide</span></h2><h3 id="model-structure-91dldxk82">📁 model Structure</h3><pre><code>📂 ComfyUI/
├── 📂 models/
│   ├── 📂 text_encoders/
│   │   ├── 📄 clip_l.safetensors
│   │   ├── 📄 t5xxl_fp16.safetensors
│   │   ├── 📄 t5xxl_fp8_e4m3fn.safetensors
│   │   └── 📄 t5xxl_fp8_e4m3fn_scaled.safetensors
│   ├── 📂 vae/
│   │   └── 📄 ae.safetensors 
│   ├── 📂 diffusion_models/
│   │   ├── 📄 fluxfill-dev.safetensors
│   └── 📂 lora/
│       └── 📄 [ICEdit-MoE.safetensors</code></pre><h2 id="essential-components-heuf1690s"><span style="color:rgb(255, 0, 242)">💎 Essential Components</span></h2><h3 id="text-encoders-the-brain-behind-natural-language-understanding-q59f65hf8">🔤 Text Encoders - The Brain Behind Natural Language Understanding</h3><ul><li><p><a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/stable-diffusion-3.5-fp8/blob/main/text_encoders/clip_l.safetensors"><strong>clip_l.safetensors</strong> - Core understanding model</a></p></li><li><p><strong>T5XXL Options</strong> - Choose based on your hardware capabilities:</p><ul><li><p><a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/stable-diffusion-3.5-fp8/blob/main/text_encoders/t5xxl_fp16.safetensors"><strong>t5xxl_fp16.safetensors</strong> - Highest quality, more demanding</a></p></li><li><p><a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/stable-diffusion-3.5-fp8/blob/main/text_encoders/t5xxl_fp8_e4m3fn.safetensors"><strong>t5xxl_fp8_e4m3fn.safetensors</strong> - Balanced performance</a></p></li><li><p><a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/stable-diffusion-3.5-fp8/blob/main/text_encoders/t5xxl_fp8_e4m3fn_scaled.safetensors"><strong>t5xxl_fp8_e4m3fn_scaled.safetensors</strong> - Optimized for speed</a></p></li></ul></li></ul><h3 id="vae-the-visual-artist-2zeo62k6r">🎭 VAE - The Visual Artist</h3><ul><li><p><a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/HiDream-I1_ComfyUI/blob/main/split_files/vae/ae.safetensors"><strong>ae.safetensors</strong></a></p></li></ul><h3 id="diffusion-models-the-imaging-engine-2aoum909n">🖼️ Diffusion Models - The Imaging Engine</h3><ul><li><p><a target="_blank" rel="ugc" href="https://civarchive.com/models/1108146/fluxfill-devgguf"><strong>fluxfill-dev.safetensors</strong> - Core diffusion model</a></p></li></ul><h3 id="lora-models-the-magic-touch-7jc3noszw">✨ LoRA Models - The Magic Touch</h3><ul><li><p><a target="_blank" rel="ugc" href="https://huggingface.co/sanaka87/ICEdit-MoE-LoRA"><strong>ICEdit-MoE.safetensors</strong> - The star of the show!</a></p></li></ul><h2 id="-24o33pbf3"></h2><h2 id="example-instructions-you-can-try-ild6kw7fn"><span style="color:rgb(255, 0, 200)">🎬 Example Instructions You Can Try</span></h2><ul><li><p>"Add a majestic castle on the mountain in the background"</p></li><li><p>"Change the color of the car from red to electric blue"</p></li><li><p>"Remove the person from the image and replace with a flowering tree"</p></li><li><p>"Transform the style to watercolor painting while maintaining all details"</p></li><li><p>"Add dramatic sunset lighting to the scene"</p></li></ul><h2 id="community-recognition-xr0mpzgmq"><span style="color:rgb(255, 0, 200)">🙏 Community Recognition</span></h2><p>A special thank you to <a target="_blank" rel="ugc" href="https://huggingface.co/sanaka87">sanaka87</a> for revolutionizing open-source image editing by releasing the ICEdit-MoE-LoRA checkpoint and demos. Your contribution empowers creators everywhere and accelerates innovation across the AI art community!</p><p>My Node OLLAMAGEMINI</p><p><a target="_blank" rel="ugc" href="https://github.com/al-swaiti/ComfyUI-OllamaGemini">https://github.com/al-swaiti/ComfyUI-OllamaGemini</a></p><h2 id="-td7naazew"><img src="https://media.licdn.com/dms/image/v2/D4E22AQGWF0ZHIpw_Cw/feedshare-shrink_1280/B4EZbKjspeHcAo-/0/1747155079532?e=1750291200&amp;v=beta&amp;t=T8ORFEFvxKcyrqVGHjDH7ms7GY-eNXoRvpuO1qEVA1E" alt="No alternative text description for this image" /></h2>

ICEdit

ICEdit-MoE-LoRA.safetensors

v1.0

<p>We present In-Context Edit, a novel approach that achieves state-of-the-art instruction-based editing <strong>using just 0.5% of the training data and 1% of the parameters required by prior SOTA methods</strong>. The first row illustrates a series of multi-turn edits, executed with high precision, while the second and third rows highlight diverse, visually impressive single-turn editing results from our method.</p><p>:open_book: For more visual results, go checkout our <a target="_blank" rel="nofollow" href="https://river-zhang.github.io/ICEdit-gh-pages/"><strong><u>project page</u></strong></a></p><p>This repository will contain the official implementation of <em>ICEdit</em>.</p><p></p><p><span><strong>⚠️ Tips</strong></span></p><p><span><strong>If you encounter such a failure case, please try again with a different seed!</strong></span></p><ul><li><p>Our base model, FLUX, does not inherently support a wide range of styles, so a large portion of our dataset involves style transfer. As a result, the model <strong>may sometimes inexplicably change your artistic style</strong>.</p><p></p></li><li><p>Our training dataset is <strong>mostly targeted at realistic images</strong>. For non-realistic images, such as <strong>anime</strong> or <strong>blurry pictures</strong>, the success rate of the editing <strong>drop and could potentially affect the final image quality</strong>.</p><p></p></li><li><p>While the success rates for adding objects, modifying color attributes, applying style transfer, and changing backgrounds are high, the success rate for object removal is relatively lower due to the low quality of the OmniEdit removal dataset.</p></li></ul><p></p><p>The current model is the one used in the experiments in the paper, trained with only 4 A800 GPUs (total batch_size = 2 x 2 x 4 = 16). In the future, we will enhance the dataset, and do scale-up, finally release a more powerful model.</p>

ICEdit-MoE-LoRA 

The current model is the one used in the experiments in the paper, trained with only 4 A800 GPUs (total batch_size = 2 x 2 x 4 = 16). In the future, we will enhance the dataset, and do scale-up, finally release a more powerful model.

Files with SHA256: aceb90fe0b20de5240ce3a5c1ab4ab7cbb819591803d98b7c41de6669ff16b18