AddModelAsLayer

Building a network piece by piece in ComfyUI is nice, but sometimes you want to build two halves separately - a feature extractor here, a classifier head there - and snap them together. AddModelAsLayer is the node for that: it takes one model (model_main) and appends every layer of a second model (model_addition) onto the end of it, returning the merged nn.Sequential. Think of it as nn.Sequential(model_main, model_addition) expressed in the graph, except it mutates model_main in place rather than building a fresh container.

You'll reach for it when a workflow benefits from two independently-constructed sub-networks. The classic pattern: build and even freeze-train a feature extractor, then attach a fresh classifier head via this node, and train the whole thing. The pack's ExtractLayersAsModel produces exactly the kind of sub-model you'd want to graft back on this way.

How it works

Both inputs must be nn.Sequential models - it raises a TypeError otherwise. It then iterates model_addition.children() and adds each one to model_main via add_module, naming them added_layer_0, added_layer_1, and so on. Your original model_main object is returned, now longer. The layers themselves are shared references, not copies - that's usually what you want when you're composing, but know that training one model trains the same parameter objects.

Inputs and output

• model_main (TORCH_MODEL) - the model being extended. The output of SequentialModelProvider or anything this pack builds.
• model_addition (TORCH_MODEL) - the sub-model whose layers get appended.

Output: one TORCH_MODEL - your model_main, extended.

Install

ComfyUI Manager, search "EternalKernel PyTorch Nodes", or:

cd ComfyUI/custom_nodes
git clone https://github.com/TashaSkyUp/EternalKernelPytorchNodes
cd EternalKernelPytorchNodes
pip install -r requirements.txt

Restart ComfyUI; node under ETK/pytorch. No model downloads - both inputs are built in-graph. Requirements are the standard stack plus scipy, scikit-learn, transformers, einops.

Common issues

• Both sides must be nn.Sequential. A bare nn.Module (say, a single layer someone passed through) raises immediately. Keep both inputs produced by SequentialModelProvider or pack builder nodes.
• Shape continuity is on you. The node glues layers end-to-end and never checks that model_main's last output dimension matches model_addition's first input dimension. A mismatch fails later at inference/training with a shape error - count your features.
• It mutates model_main in place. If you wired model_main elsewhere in the graph expecting the pre-merge version, it's now longer. Duplicate before merging if you need both.
• Shared parameters. Because layers are shared references, training after a merge trains the originals too. For a "freeze the backbone" workflow you'd still want freeze=True on the extractor side or SetModelTrainable afterward.

No community tutorials, no threads - this pack is a quiet corner. The upside: it's two torch primitives, so behavior is predictable from PyTorch docs. Pack-wide quirk to remember: it patches ComfyUI's validator to ignore return_type_mismatch, so a wrong wire might not produce the error you're hunting.