VRAM Gated LoRA Loader (Model Only)

The VRAM Gated LoRA Loader (Model Only) is the fourth and final member of TrentNodes' gated family, and it solves a specific sequencing annoyance: you have a caption-then-generate graph, the vision model needs to finish and free VRAM first, and your LoRA shouldn't be applied until the big diffusion model is actually in memory. So this loader waits for VidScribe's vram_cleared signal before it does its thing - same pattern as the gated checkpoint, VAE, and diffusion-model loaders, just for LoRAs.

"Model Only" in the name means exactly what it says: this node applies LoRA weights to a MODEL and only a model. There's no CLIP output, no strength_clip slider - if you're loading a text-side LoRA that changes prompt behavior, you want a full LoRA loader, not this one. This is the variant you chain after a gated UNET or checkpoint loader when the LoRA is the last thing before the sampler.

Inputs and outputs

• vram_signal - wire from VidScribe's vram_cleared. The required-input gating is the whole mechanism: the node can't run before the signal exists, so ComfyUI orders it after the VLM. It's execution sequencing, not a VRAM meter.
• model - the MODEL to patch. Feed it the gated UNET loader's output.
• lora_name - dropdown from ComfyUI/models/loras.
• strength_model - default 1.0, range −100 to 100. Set to 0 and the node returns the model unchanged (no-op), which is handy for A/B testing without unplugging the node.

Output: a single model with the LoRA baked in.

Where it fits

The full ladder in one of these graphs looks like: VidScribe describes → vram_cleared fans out to the gated checkpoint/UNET loader and the gated LoRA loader → model flows through LoRA → sampler. Every gated node in the family shares the same signal, so wiring them in parallel from one output is the intended use, not a hack.

Troubleshooting

Two beginner traps. First, forgetting the model input: this loader needs an actual model to patch, and unlike a checkpoint loader it doesn't bring its own. Second, expecting CLIP-side behavior: no strength_clip here, so a LoRA that was trained with both model and CLIP weights only gets its model half applied. If your output looks texturally off rather than prompt-wise, that's usually why.

Installing

Ships with TrentNodes under Trent/VLM. ComfyUI Manager → "Trent Nodes" → Install, or clone it:

cd ComfyUI/custom_nodes
git clone https://github.com/TrentHunter82/TrentNodes
cd TrentNodes
pip install -r requirements.txt

Restart after. If Manager flags the pack (a day-one repo rename left a registry duplicate it still distrusts), the manual clone above is the reliable path.