Load Model Patch

This is the node ComfyUI grew when the newest control techniques stopped fitting the old loader slots. Load Model Patch reads a file from ComfyUI/models/model_patches and hands you a single MODEL_PATCH wire. Sounds unremarkable. What's actually going on is that one node now auto-detects and loads a grab bag of the most recent conditioning tech - Anima's LLLite, Qwen-Image's block-wise ControlNet, Z-Image's Fun ControlNet, Wan's Uni3C, even a SUPIR-style denoiser - all through the same interface. It's marked experimental in the source, and it deserves the label: this is a loader for people who follow model releases weekly, not for the Monday-morning workflow.

How it works

One input, name, listing whatever .safetensors files you've dropped in models/model_patches. On load it sniffs the weights and picks the right model class - the same trick CheckpointLoaderSimple uses to guess architectures, but tuned for patches:

• lllite_conditioning1.conv1.weight → an Anima LLLite (kohya's lightweight control-adapter line)
• controlnet_blocks.0.y_rms.weight → Qwen-Image block-wise ControlNet
• control_all_x_embedder keys → the Z-Image Fun ControlNet family (including a broken flag it sets when it detects a refiner with zeroed weights)
• controlnet_patch_embedding.weight → Wan Uni3C, the point-cloud/human-model control net
• audio_proj.proj1.weight → a MultiTalk audio patch
• first_stage_model.denoise_encoder keys → a SUPIR denoiser

What you do with the resulting MODEL_PATCH depends on which it turned out to be. It's not a universal "apply" wire. Anima LLLite patches go into Anima LLLite Apply; Qwen and Z-Image ones into the Qwen Image Diffsynth ControlNet / Z-Image Fun ControlNet apply nodes (which take the base model, VAE, and a conditioning image plus strength); Uni3C goes into Wan Uni3C ControlNet Apply with a render video; the SUPIR one is for restoration. In every case the pattern is: loader gives you the patch, an apply node fuses it into your base MODEL.

Where it came from, and whether you need it

This node is a 2025–2026 addition, and it's the direct answer to a real problem in the control landscape. Traditional ControlNet - a trained copy of the encoder - was never rebuilt for every new DiT architecture; instead the community shipped lighter patches (LLLite, Uni3C, Fun ControlNet) that inject control at attention and MLP points. ComfyUI has been folding those under one loader rather than one node per technique. Community signal is thin but telling: Uni3C threads are about wanting it ("Uni3C camera moves applied to still images for WAN 2.2 anywhere?") more than celebrating it, and the "model patch" name only really shows up in the corpus from late 2025, which tracks this node landing.

So, honestly: if you're not actively running Anima, Qwen-Image, Z-Image, or Wan control workflows, you won't touch this node - most of those apply nodes are also experimental and clearly labeled. When you do need it, the main practical gotchas are:

• Right file, wrong folder. It's model_patches, not controlnet. Easy miss, since most of these files are control nets in spirit.
• The loader can't tell you what it found. If a format it doesn't recognize shows up, you get an error rather than a hint. Check the file's source repo if a load fails.
• Treat "experimental" as a warning, not a meme. These patches are where the bleeding edge lives; versions change and workflows break between ComfyUI updates.

It ships with ComfyUI core - no install. Just don't expect it in any of the beginner default templates. Those don't need it. You will, eventually.