ComfyUI - Jurdns Model Sculptor

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.

What Does It Do?

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.

Use Cases

🔍 Iterative Upscaling & Refining

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.

🎨 Creative Model Merging Process

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.

⚡ Dynamic Workflow Enhancement

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.

Installation

1. Navigate to your ComfyUI custom nodes directory:

   cd ComfyUI/custom_nodes/

2. Clone this repository:

   git clone https://github.com/jurdnf/ComfyUI-JurdnsModelSculptor.git

Also found in the ComfyUI Manager.

This post also contains a .zip archive that you can extract directly into comfyui/custom_nodes

1. Restart ComfyUI

The nodes will appear under models/advanced in your node browser.

Available Nodes

• Jurdn's Model Sculptor (Flux) - For Flux.1 models

• Jurdn's Model Sculptor (SDXL) - For SDXL models

• Jurdn's Model Sculptor (SD3) - For Stable Diffusion 3 models

Finetuned versions of the above architectures (such as Illustrious, because it is based on SDXL)

How to Use

1. Connect your model from "Load Diffusion Model" directly to a Model Sculptor node

2. Choose a gradient shape that defines how the effect varies across layers

3. Set the strength (typically 0.05-0.3 for subtle effects, higher for dramatic changes)

4. Select target blocks to focus the effect on specific parts of the model

5. Connect the output to your sampler as usual

Important: Always connect directly from the model loader to avoid cumulative effects.

Gradient Shapes Explained

Each shape creates a different "curve" of influence across your model's layers:

• Linear (Ascending/Descending) - Gradual ramp up or down

• Ease In/Out (Quadratic) - Smooth acceleration/deceleration curves

• Ease In/Out (Sine) - Natural, smooth wave-like transitions

• Spike (Gaussian) - Sharp peak in the middle layers

• Dip (Inverse Gaussian) - Valley in the middle, emphasis on edges

• Steps - Discrete level changes across layers

• Random (Noise) - Unpredictable variations for experimental effects

Target Blocks

Flux Models

• in_layers - Input processing layers

• double_blocks - Main transformer blocks (0-18)

• single_blocks - Secondary transformer blocks (0-37)

• Double & Single (Synced Shape) - Apply same gradient to both block types

SDXL Models

• input_blocks - Downsampling path (0-11)

• middle_block - Bottleneck processing

• output_blocks - Upsampling path (0-11)

• time_embed - Timestep embeddings

• label_emb - Class/style embeddings

• Input & Output (Synced Shape) - Symmetrical application

SD3 Models

• joint_blocks - Main DiT transformer blocks (0-23)

• x_embedder - Image embedding layers

• y_embedder - Text embedding layers

• t_embedder - Time embedding layers

• pos_embed - Positional embeddings

• final_layer - Output projection

• Joint & Final (Synced Shape) - Focus on core processing

Tips & Best Practices

• Start subtle - Begin with strength values around 0.1 and adjust

• Experiment with combinations - Different shapes work better for different content

• Use in upscaling - Particularly effective during multi-stage upscaling workflows

• Connect directly - Always connect from "Load Diffusion Model" to avoid stacking effects

• Try different blocks - Each block type affects different aspects of generation