SID LLM Local

Every cloud provider on the toolkit's other node wants a key and a credit card. This one wants neither. SID_LLM_Local is the pack's free option: a vision LLM that runs inside your ComfyUI process, with the weights pulled from Hugging Face on first use. The node description says "No GPU" - meaning it'll happily run on CPU if that's all you've got (slow, but it runs) - and no API. For a node nobody's searching for yet, it's quietly the one most people should start with.

Which model? The dropdown lists Qwen3-VL (2B / 4B / 8B, plus Thinking reasoning variants), Florence-2, Moondream2, SmolVLM, and Phi-3.5-Vision. The labels carry the spec sheet: max tokens and VRAM at 4-bit. The 8B sits at ~4.5GB quantized but wants 12GB at full fp16, which is why the pack defaults to the 2B. Community consensus from the release thread: 4B is the sweet spot, "you might even get away with the 2B."

How it works. It loads the model with Hugging Face transformers in-process - no Ollama process to babysit - auto-detects CUDA vs CPU, and quantizes to fit your card. quantization handles that (Auto / 4-bit / FP16). keep_model_loaded keeps it warm in VRAM between runs, which turns repeat inference from a multi-second load into instant. The Thinking variants set the supports_reasoning flag that flips the prompt generator into agentic mode; the plain variants stay single-shot, which the pack's own docs recommend for local models anyway.

Inputs. model, quantization, device, temperature, max_tokens_preset, custom_max_tokens, keep_model_loaded, attention_mode, repetition_penalty, top_p, use_torch_compile. You realistically touch three:

• model - the vision model. Pick the smallest that gives you usable prompts; 4B Qwen3-VL is the "good enough" default for most people.
• quantization - Auto is fine; drop to 4-bit only if you're VRAM-starved.
• keep_model_loaded - leave it on unless you need the VRAM back between runs.

The rest are pass-through sampling knobs: repetition_penalty (1.2 default) is worth knowing if output starts looping, and attention_mode can go flash_attention_2 on an Ampere+ card for a speed bump. use_torch_compile speeds inference but makes the first run slower and needs CUDA + Torch 2.1+.

Output. One LLM_MODEL, wired straight into the llm_model input of SID_ZImagePromptGenerator. Same contract as the API node, so swapping between the two is a two-second edit.

Install. ComfyUI Manager (search "AI Photography Toolkit"), or:

cd ComfyUI/custom_nodes
git clone https://github.com/slahiri/ComfyUI-AI-Photography-Toolkit.git

Restart ComfyUI; the pack installs transformers, accelerate, and friends automatically. 4-bit quantization needs bitsandbytes and is CUDA-only.

Gotchas. The first run downloads multiple GB of weights, so the first generation looks hung - it isn't. If you're on 8GB of VRAM, don't reach for the 8B at fp16; the 4-bit 4B will make you happier. And if local output comes out rambly or structured-JSON-shaped, you picked a Thinking variant - switch to a plain one and let it run single-shot, which is where small local models shine.