🧬 Gimbal Likeness Isolator

Character LoRAs have a chronic problem: crank the strength to nail the face, and the background styling, clothes, even the pose get dragged along with it. Gimbal Likeness Isolator is the pack's answer - a LoRA loader that splits the knob in two, so you can push the identity without pushing everything else. It's the node in the Gimbal-comfy suite (by Form & Noise) that reads "if you already use LoRAs, this is the one that needs no latent math."

Let's be honest about what it is, because the marketing oversells: this is not a model that has learned to separate identity from style. It's a wrapper around ComfyUI's standard load_lora_for_models that gives you three knobs instead of one, and the trick is where each knob lands.

How it works

The mechanism, from the source: it loads the LoRA file from your loras folder and patches both the model (UNet) and clip with it. The three inputs map to two actual patch strengths:

• strength - the model patch strength. This drives the visual/UNet side, the part that changes pose, materials, lighting structure.
• alpha - the base CLIP strength.
• likeness_mask - a multiplier on the CLIP side. The effective CLIP strength becomes alpha × likeness_mask.

So "likeness" here means the text-encoder contribution of the LoRA - the tokens that carry the identity description. Push likeness_mask up to 3 and you're weighting those identity tokens heavily in the text encoder while leaving the UNet patch at whatever strength says. Drop it toward 0 and you keep the LoRA's stylistic influence while gutting its identity. Negative strengths invert the LoRA's direction - useful for pushing away from a character, which is a legitimate "vector" move even if it's rarely what you want on a first try.

The honest catch: whether this actually separates identity from scene depends entirely on how the LoRA was trained. A well-captioned character LoRA (identity in the captions, style left implicit - see the KB's lora-training notes on this) responds beautifully to CLIP-side weighting. A LoRA with style baked into the training data won't be separable, because the two are entangled in the weights themselves. No loader can un-entangle that.

The inputs that matter

lora_name is a dropdown populated from ComfyUI/models/loras/ at runtime - it shows "<no loras found>" until you have files there. Then it's just strength and likeness_mask; leave alpha at 1 until you want to scale the whole CLIP side. If strength, alpha, and likeness_mask are all at neutral (0, 0, 1.0), the node short-circuits and passes the model and clip through untouched - a no-op, so you can leave it in the graph without side effects.

Outputs are the patched MODEL and CLIP, wired exactly like any LoRA loader: MODEL into the KSampler, CLIP into your CLIP Text Encode.

Install

# ComfyUI Manager → search "Gimbal"
# or:
cd ComfyUI/custom_nodes
git clone https://github.com/FormAndNoise/Gimbal-comfy
pip install -r Gimbal-comfy/requirements.txt

Restart ComfyUI; it's under Add Node → Gimbal/Flight Instruments as "🧬 Gimbal Likeness Isolator". Dependencies are just torch/numpy/pillow - no downloads beyond whatever LoRAs you already have.