<h1 id="comfyui-jurdns-model-sculptor-vyyxymco5">ComfyUI - Jurdns Model Sculptor</h1>A collection of ComfyUI nodes that "sculpt" diffusion models by applying gradient-based modifications to different layers and blocks. Transform your models with mathematical curves and shapes to create unique artistic effects, enhance details during upscaling, or experiment with creative model variations.<h2 id="what-does-it-do-rnxxanazr">What Does It Do?</h2>Instead of traditional model merging that blends two models together, Model Sculptor applies mathematical gradient shapes across the layers of a single model. Think of it like applying a curve or wave pattern to the model's "depth" - making some layers stronger, others weaker, creating peaks and valleys of influence throughout the network. Best used in refinement or upscale stages for effective detailing.<h2 id="use-cases-7ulq3of3d">Use Cases</h2><h3 id="iterative-upscaling-and-refining-0bzsiqwae">🔍 Iterative Upscaling &amp; Refining</h3>Use Model Sculptor in your upscaling workflows to enhance detail recovery. Apply different gradient shapes at different upscale stages to bring out fine details, textures, or specific artistic elements.<h3 id="creative-model-merging-process-g4r1r8af9">🎨 Creative Model Merging Process</h3>Instead of traditional merging, use Model Sculptor as a preprocessing step before merging, or apply it to merged models to create unique variations that can't be achieved through conventional blending.<h3 id="dynamic-workflow-enhancement-ythecvwsi">⚡ Dynamic Workflow Enhancement</h3>Integrate into your existing workflows where you want the same base model to behave differently for specific tasks - portraits vs landscapes, detailed vs stylized, etc.<h2 id="installation-ziq21ieyw">Installation</h2><ol><li>Navigate to your ComfyUI custom nodes directory:<pre><code>cd ComfyUI/custom_nodes/</code></pre></li><li>Clone this repository:<pre><code>git clone https://github.com/jurdnf/ComfyUI-JurdnsModelSculptor.git</code></pre></li></ol>Also found in the ComfyUI Manager.This post also contains a .zip archive that you can extract directly into comfyui/custom_nodes<ol><li>Restart ComfyUI</li></ol>The nodes will appear under models/advanced in your node browser.<h2 id="available-nodes-dhzk8y6q6">Available Nodes</h2><ul><li>Jurdn's Model Sculptor (Flux) - For Flux.1 models</li><li>Jurdn's Model Sculptor (SDXL) - For SDXL models</li><li>Jurdn's Model Sculptor (SD3) - For Stable Diffusion 3 models</li></ul>Finetuned versions of the above architectures (such as Illustrious, because it is based on SDXL)<h2 id="how-to-use-u9kp5izcf">How to Use</h2><ol><li>Connect your model from "Load Diffusion Model" directly to a Model Sculptor node</li><li>Choose a gradient shape that defines how the effect varies across layers</li><li>Set the strength (typically 0.05-0.3 for subtle effects, higher for dramatic changes)</li><li>Select target blocks to focus the effect on specific parts of the model</li><li>Connect the output to your sampler as usual</li></ol>Important: Always connect directly from the model loader to avoid cumulative effects.<h2 id="gradient-shapes-explained-s86f86hfo">Gradient Shapes Explained</h2>Each shape creates a different "curve" of influence across your model's layers:<ul><li>Linear (Ascending/Descending) - Gradual ramp up or down</li><li>Ease In/Out (Quadratic) - Smooth acceleration/deceleration curves</li><li>Ease In/Out (Sine) - Natural, smooth wave-like transitions</li><li>Spike (Gaussian) - Sharp peak in the middle layers</li><li>Dip (Inverse Gaussian) - Valley in the middle, emphasis on edges</li><li>Steps - Discrete level changes across layers</li><li>Random (Noise) - Unpredictable variations for experimental effects</li></ul><h2 id="target-blocks-b4m797j99">Target Blocks</h2><h3 id="flux-models-0lcd6cqbv">Flux Models</h3><ul><li>in_layers - Input processing layers</li><li>double_blocks - Main transformer blocks (0-18)</li><li>single_blocks - Secondary transformer blocks (0-37)</li><li>Double &amp; Single (Synced Shape) - Apply same gradient to both block types</li></ul><h3 id="sdxl-models-sa2c99wiy">SDXL Models</h3><ul><li>input_blocks - Downsampling path (0-11)</li><li>middle_block - Bottleneck processing</li><li>output_blocks - Upsampling path (0-11)</li><li>time_embed - Timestep embeddings</li><li>label_emb - Class/style embeddings</li><li>Input &amp; Output (Synced Shape) - Symmetrical application</li></ul><h3 id="sd3-models-hs8bdtlxf">SD3 Models</h3><ul><li>joint_blocks - Main DiT transformer blocks (0-23)</li><li>x_embedder - Image embedding layers</li><li>y_embedder - Text embedding layers</li><li>t_embedder - Time embedding layers</li><li>pos_embed - Positional embeddings</li><li>final_layer - Output projection</li><li>Joint &amp; Final (Synced Shape) - Focus on core processing</li></ul><h2 id="tips-and-best-practices-o1vrpgjsm">Tips &amp; Best Practices</h2><ul><li>Start subtle - Begin with strength values around 0.1 and adjust</li><li>Experiment with combinations - Different shapes work better for different content</li><li>Use in upscaling - Particularly effective during multi-stage upscaling workflows</li><li>Connect directly - Always connect from "Load Diffusion Model" to avoid stacking effects</li><li>Try different blocks - Each block type affects different aspects of generation</li></ul>

Files with SHA256: 058d2ca239739c3332b014e4f71f7d00533929250852fa766435aa1c9220a0cf