TS Qwen 3

The name says it all: it's a Qwen 3 VL vision-language model, running locally, as a ComfyUI node. No API, no key, no per-token bill. Give it an image (or a video, or just text) and a question, and it answers - which in a ComfyUI graph means captioning training data, describing what's in a frame, or turning a rough idea into a proper prompt for the next sampler down the line. It's the same category of tool the KB calls the local prompt-enhancement worker: uncensored, offline, and free per call.

How it works

Under the hood it's a shared Qwen engine (nodes/llm/_qwen_engine.py), the same one TS Super Prompt's voice-enhancement uses - so bug fixes and performance improvements land in both nodes at once. The model downloads from Hugging Face on first use into models/LLM/, and the built-in picker offers 2B / 4B / 8B variants plus uncensored ("abliterated") builds - the default model is the huihui-ai Qwen3.5-2B abliterated, which is a sensible starting point for a 2B worker. Qwen3-VL is Apache 2.0, and as the KB's Qwen3-VL panel notes, this community's whole reason for running the abliterated builds is that the stock Instruct models carry assistant-style refusals that break explicit captioning.

The engine handles image, video, and text in one node: image takes an IMAGE, video takes a frame batch (with video_max_frames, default 16, controlling how much of the clip the model sees), and prompt is your question.

The inputs that matter

• system_preset - behavior presets like "Prompts enhance" and "Image Edit Command Translation" (pick "Your instruction" and fill custom_system_prompt for your own).
• precision - auto by default. int8/int4 options only appear if bitsandbytes is installed - that's the VRAM-saver that lets an 8B run on a small card.
• attention_mode - flash_attention_2 is faster and leaner on compatible GPUs.
• max_new_tokens - response length (512 default).
• seed - reproducible output (default 42).
• unload_after_generation - free the model's VRAM right after each run. Saves memory, slows repeated runs.

Two outputs: generated_text (the answer, as a STRING you can feed anywhere) and processed_image (the input image passed through unchanged - handy when the node sits in a chain). enable (default true) lets you bypass the model entirely for a pure pass-through.

The security and token warnings

This is a node that downloads from the internet and runs arbitrary local weights - the same shape as the pack that got ComfyUI in trouble, so the KB's advice applies: only install it because you trust the pack (it's a mature, well-reviewed repo, not a random fork). Two specific gotchas:

• hf_token is saved in the workflow JSON in plain text. The tooltip warns you directly: don't share a file that contains a token. Leave it empty for public models - almost everything here is public.
• offline_mode exists precisely for machines that shouldn't phone home: with it on, the node uses only files already in models/LLM/.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/AlexYez/comfyui-timesaver
cd comfyui-timesaver
python -m pip install -r requirements.txt

The pack requires transformers >= 5.2.0 (its default model's architecture is only understood from that version - older installs downloaded gigabytes before failing, now it checks first). For int4/int8 precision:

python -m pip install -e .[llm-quant]

Note: bitsandbytes has no Apple Silicon wheel, so on a Mac the node silently falls back to fp16/fp32 with a warning - that's expected, not a bug.

Common issues

• OOM on an 8B. Drop precision to int8 or int4, set unload_after_generation on, and cap max_image_size / video_max_frames.
• "Transformers does not recognize this architecture." Update transformers to 5.2.0+ - the pack now checks this before downloading, so you'll get a clean error with the fix.
• Slow first run. That's the model downloading into models/LLM/, not a hang.

For captioning a dataset for LoRA training, or routing "user intent → edit command" into a Qwen-Edit graph, this is the local worker to wire in.