TS Super Prompt

The blank prompt box is the real enemy. TS Super Prompt is the pack's answer to it, and it's the friendliest version of this idea going: it has a built-in voice button. Click the mic, describe your shot out loud, and Whisper transcribes it - with cinematography-aware grammar fixes - then a small local Qwen3 model expands it into a rich, structured prompt. No API key, no cloud, everything on your machine.

That's the headline, but the node earns its keep in the details, and there are a lot of them.

How it works

The flow is two local models working in sequence. Whisper turns your speech into text (the high_quality toggle picks the faster large-v3-turbo over the quick base model). Then a Qwen3.5 model - 2B by default, or 4B if you flip bigger_model - rewrites the text according to a system_preset chosen from the pack's qwen_3_vl_presets.json. The enhanced text replaces the field and flows out the single text output, ready for a CLIPTextEncode.

Two things about the download path, because this is where LLM nodes usually burn people:

• The model is never downloaded silently. If it's not on the machine, pressing the enhance button opens a dialog first: which model, exactly how large (the real repo size, asked of the Hugging Face API without fetching a byte), and which folder it lands in. No click, no download.
• The library is checked before the download. The node asks the hub what architecture the model is and compares it with what your installed transformers understands, and stops immediately if it can't load it - instead of spending minutes and gigabytes to reach the same conclusion. The default Qwen3.5 model needs transformers 5.2.0 or newer, which is what the pack's requirements.txt pins.

The reference image part

You can attach a reference image right in the node - drag it in, paste it, pick it. One image is a plain reference. Drop a second and the two become the first and last frame of the shot, and the model is told explicitly. There's also an optional images input for frames that come from the graph itself: in a batch, the first image is the first frame and the last is the last frame. A wired input wins over attached images, as a whole.

Here's the clever bit: the Enhance button sees the input too. It only computes the branch of the graph that feeds this input - pulled out into its own prompt, run on its own - so no sampler or save node fires just because you hit enhance. And on-screen text is not translated: anything in quotes (" ", « », ' ') is copied through character-for-character in its original language, which is what you want when a sign or a lyric must appear exactly as written.

There are video presets too (Video Prompt Enhance LTX for LTX-2.5, Video Prompt Enhance H3 for MiniMax H3) written to each model's own prompt rules, including H3's dialogue schema for spoken Russian that actually sounds Russian.

The controls that matter

• text - the prompt field itself; recognized speech lands here and the enhance button replaces it.
• system_preset - which enhancement behavior to use.
• bigger_model - off is the 2B (fast, already downloaded); on is the 4B (better prompts, about twice the download and twice the VRAM, fetched on first use).
• attached_image / attached_image_2 - internal fields the Attach button fills; you interact with the images, not these.

Installing it

Part of comfyui-timesaver: ComfyUI Manager → search Timesaver, or

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

then restart ComfyUI. Whisper and the LLM stack are in the base requirements, so a plain install covers everything.

Gotchas

A local 2B model is a convenience, not a writer - it will cheerfully add detail you didn't ask for, same as any enhancer. Read the enhanced prompt before you trust it, especially for anything with a specific subject. And if the enhance button says the model can't load, it's almost always the transformers version: check it with python -m pip show transformers, and re-run the pack's requirements.txt if it's below 5.2.0.