Jurdn's Model Sculptor (SD3)

Let's be honest up front: this is the least useful node of the three in Jurdn's Model Sculptor pack, and it's not the node's fault. It sculpts Stable Diffusion 3 models - and SD3's ecosystem never formed. The June 2024 launch was such a mess it got banned from CivitAI, SD 3.5 was genuinely competent but almost nobody built on it, and the field moved to Flux and its successors within months. So the only reason you'd be here is if you already run SD3 or SD3.5 checkpoints. If you do, this works exactly like its SDXL and Flux siblings: a per-layer weight scaler that reshapes a loaded model without training, merging, or touching any file.

How it works

Same mechanism as the rest of the pack, mapped onto SD3's MMDiT transformer. The node clones the loaded model, pulls the weights under diffusion_model., generates a curve with one value per selected component, and applies each value as a weight scale. Through ComfyUI's patch system each affected layer's weights end up multiplied by roughly 1 + (curve value × strength). Default strength 0.1 means a Spike (Gaussian) lifts the middle joint blocks by about 10% and leaves the ends alone. Negative strength flips the curve. Nothing is saved to disk, and the original model file is never written.

The inputs that matter

Four required inputs, same shape as its siblings:

• model - wire it straight from Load Diffusion Model.
• gradient_shape - ten curves: Linear up/down, Ease In/Out (Quadratic and Sine), Spike (Gaussian), Dip (Inverse Gaussian), Steps up/down, Random (Noise).
• strength - default 0.1, range −2 to 2. A multiplier, so keep it in the 0.05–0.3 band for subtle work.
• target_blocks - eight options. joint_blocks 0–23 are the 24 main MMDiT blocks (the real meat), x_embedder handles image tokens, y_embedder the text tokens, t_embedder timesteps, pos_embed positions, final_layer the output projection, plus "all" and Joint & Final (Synced Shape).

The interesting knob for SD3 specifically is y_embedder - scaling the text-embedding layers is a cheap proxy for conditioning strength, and it does more than you'd expect at small values. Output is a single MODEL that replaces the loader's connection into your sampler.

Install

Same pack as the Flux and SDXL sculptors, so installing it installs all three:

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

Restart ComfyUI afterwards, or search ComfyUI Manager for "ComfyUI-JurdnsModelSculptor". There's no requirements.txt - no extra pip packages, no model downloads. The nodes appear under models/advanced.

Where people get burned

• Architecture mismatch. If the node can't find matching layer names, the console prints Found 0 patches matching target prefixes and it hands your model back unchanged. The author targets SD3; SD 3.5 is a different architecture with a different block layout, so if you feed in an SD3.5 checkpoint and nothing changes, this is why. Check the terminal.
• Stacking. Connect directly from Load Diffusion Model - the node patches a clone, and chaining it after other patching nodes compounds effects you didn't intend.
• Random (Noise). Unseeded RNG means a new sculpt every run. Fine for poking around, useless for reproducibility. Use a deterministic shape.

Real talk: unless you're an SD3 holdout or testing the MMDiT architecture, the Flux and SDXL siblings are where this pack's value lives. This one costs nothing extra to have installed, though - the whole pack is one clone and zero dependencies, so keep it for the day you actually load an SD3 model.