Apply Null-LaLoRA (Universal + Null)

ApplyNullLaLoRA (displayed as "Apply Null-LaLoRA (Universal + Null)") is the most research-forward node in ComfyUI_UniversalSubspace, and the one you'll use last - if you use it at all. The others handle "plain" universal weights; this one applies a specific training scheme the author calls Null-LaLoRA, built to combine a frozen null-space component with a trainable subspace component.

The idea

The display name says it: "universal + null." A Null-LaLoRA splits each layer's adapter in two:

• a frozen null part - weights pinned so they don't disturb what the base model already knows, which is the trick for avoiding forgetting or collisions when you stack adapters;
• a trainable universal part - coordinates along the shared subspace basis (the same μ + V·α scheme as the rest of the pack), which carries the new behavior the training was actually about.

Optionally it adds DoRA-style magnitude scaling on top. The docstring names the companion training script, network_null_lalora_bata.py - note that this script is not in the pack. The repo only ships inference; the training half lives in the author's research pipeline.

What the node does

It reconstructs each layer's down/up matrices, concatenates the frozen and trainable pieces along the rank dimension, applies either standard (s) or DoRA (m) scaling, and hands the assembled LoRA to ComfyUI's load_lora_for_models like any other apply node. Two details worth knowing:

• It deliberately processes the reconstruction on CPU to save VRAM. Your GPU stays free during the math, at the cost of speed - for big LoRAs the apply can take a beat.
• The weights file has to carry the null parts, either saved inside it (keys like .null_up / .null_down) or pulled from the subspace file itself. If neither exists for a layer, the layer is silently skipped.

The inputs

• model, clip - as always.
• subspace - UNIVERSAL_SUBSPACE from UniversalLoRALoader.
• null_lalora_weights - a UNIVERSAL_WEIGHTS file from LoadUniversalWeights, but in the Null-LaLoRA format: alpha_down / alpha_up for the trainable part, s or m for scaling.
• strength_model / strength_clip - default 1.0, range -10 to 10.

Outputs are MODEL and CLIP, wired into sampler and conditioning as usual.

The honest take

This node only does something if you have a Null-LaLoRA trained with the author's pipeline. That's a tiny audience: people following this specific research line. If you're not one of them, this node will load fine, accept a weights file, and quietly do nothing - the console prints Null-LaLoRA Applied Layers: 0 and the model passes through untouched. Don't mistake that for a bug. It's the pack working as intended on data it can't handle. Check the console, verify your weights file actually contains alpha_down / alpha_up and null parts, and confirm the layer naming matches the basis before assuming something's broken.

Install

Identical to the rest of the pack - no extra dependencies, nothing heavy to pull:

# ComfyUI Manager: Manager → Install Custom Nodes → search "UniversalSubspace" → Install → Restart

cd ComfyUI/custom_nodes
git clone https://github.com/a-ru2016/ComfyUI_UniversalSubspace

Then restart ComfyUI. It'll appear under the UniversalSubspace category alongside the four siblings. And yes - this whole pack is research-grade with effectively no community usage, so if this node looks over your head, that's not a value judgment on you. It's a value judgment on the node.