LayerUtility: Load SmolVLM Model(Advance)

If LlamaVision elsewhere in this pack is the heavyweight option - an 11B model that needs real VRAM even quantized - this is the other end of the spectrum. SmolVLM is HuggingFace's deliberately lightweight vision-language model, and this node just loads it so a downstream captioning node (SmolVLM, not covered here but part of the same pair) can use it. Reach for this when you want quick, cheap image understanding and don't need the depth of reasoning a bigger model gives you - batch captioning, simple "what's in this image" checks, anything where speed matters more than nuance.

The mechanism is a pure loader, same pattern as LoadBiRefNetModel and LoadSAM2Model elsewhere in the pack: pick your options, it loads the checkpoint into memory, and hands it downstream as a model object rather than doing any image processing itself.

All three inputs:

• model - currently only one option, SmolVLM-Instruct. No other SmolVLM variants are exposed in this dropdown as of this pack version.
• dtype - bf16 or fp32. bf16 halves memory usage with minimal accuracy loss for inference, which is the practical default unless you have a specific reason to want full precision.
• device - cuda or cpu. SmolVLM is small enough that CPU inference is actually viable here in a way it wouldn't be for something like LlamaVision, if you're VRAM-constrained and willing to trade speed.

Output: smolVLM_model, a SmolVLM_MODEL object that wires straight into the pack's separate SmolVLM inference node.

Installing it: ComfyUI Manager, search "ComfyUI Layer Style Advance," or git clone https://github.com/chflame163/ComfyUI_LayerStyle_Advance into custom_nodes, then install_requirements.bat (or Aki variant) / pip install -r requirements.txt, then repair_dependency.bat, restart. The model itself needs to be downloaded separately - grab the SmolVLM-Instruct folder from huggingface.co/HuggingFaceTB/SmolVLM-Instruct (or the BaiduNetdisk mirror, bundled together with the SmolLM2 text-model downloads in the same archive) and copy it into ComfyUI/models/smol.

Where people get stuck: an empty or errored load almost always traces back to the model not actually being in ComfyUI/models/smol yet - unlike some of the pack's other model-backed nodes, there's no automatic first-use download here, you need to fetch it yourself before this loader has anything to find. Beyond that, this node shares the same general dependency footprint as the rest of the pack (this repo exists specifically to carry the nodes with heavier transformers/vision-model dependencies split off from the base LayerStyle pack), so if the node fails to import entirely rather than just failing to find a model, run repair_dependency.bat first before troubleshooting anything model-specific - that clears the more common transformers-version conflicts that affect the whole pack's import, not just this node.