Vision Prompt Assistant

Vision Prompt Assistant is the heart of this pack, and the first thing to know is that the display name is doing the heavy lifting: it's called an "assistant" but it's a local text-generation node - no API, no key, nothing leaves your machine. You drop it in front of a MiniMax H3 workflow, connect up to three reference images, and a multimodal text encoder on your GPU reads them and writes the generation prompt for you.

Why this matters for H3 specifically: MiniMax H3 is an omni-modal model that treats text, image, video and audio as one context, and it follows a well-written prompt remarkably closely - the launch-day community reaction was people shocked that a long, detailed prompt got followed to the letter. But writing that prompt by hand is the bottleneck. This node removes the blank page by turning reference images into Picture-tagged prose that maps directly to H3's reference syntax.

How it works

It loads a multimodal CLIP through ComfyUI's own text-encoder loader - the author's recommended build is qwen3-vl-4b-heretic_int8.safetensors from DreamFast's Qwen3-VL-4b-Heretic-ComfyUI repo (the "heretic" part being an abliterated Qwen3-VL, since the stock Instruct models carry assistant-style refusals that are useless for explicit captioning work). You select it in the clip_name dropdown and start with clip_type set to ltxv, though you can compare against stable_diffusion.

The loaded encoder is cached and reused while the model, type, and device stay the same, so editing a prompt doesn't re-read the checkpoint from disk every run.

The inputs that matter

• clip_name - pick the Qwen3-VL encoder. Drop the .safetensors into ComfyUI/models/text_encoders first.
• clip_type - ltxv is the recommended starting point.
• image_0 / image_1 / image_2 - up to three references. Socket order maps to tags: first image becomes <Picture 1>, second <Picture 2>, third <Picture 3>.
• user_prompt / system_prompt - the defaults already write H3-ready prompts; user_prompt is the preferred STRING route when the node is bypassed.
• max_length - the hard cap for new tokens. The node automatically budgets slightly below it so the model can finish its final sentence instead of being chopped off.
• temperature / top_k / top_p / min_p / repetition_penalty - standard sampling; the defaults (0.7 / 40 / 0.9 / 0.05 / 1.05) are a reasonable starting set.

Output

A single generated_text string. Feed it to a Preview Vision Prompt, or straight into H3's conditioning.

Install

ComfyUI Manager → search Vision Prompt Assistant, or:

cd ComfyUI/custom_nodes
git clone https://github.com/elgalardi/ComfyUI-VisionPromptAssistant

Then restart ComfyUI. The one real download is the encoder:

# into ComfyUI/models/text_encoders
wget https://huggingface.co/DreamFast/Qwen3-VL-4b-Heretic-ComfyUI/resolve/main/qwen3-vl-4b-heretic_int8.safetensors

Getting fuller output

The author's trick, and it's a good one: end user_prompt with an explicit token target slightly below max_length. With max_length at 256, finish with "Write about 220 tokens." The model knows the budget and stops writing early - a short prompt from a 4B encoder is worse than no prompt at all, so it's worth the one line.

Two things this node deliberately does not do: it only accepts images, not video or audio - Qwen3-VL through ComfyUI's text-generation interface doesn't reliably understand a full video clip or listen to a waveform, and the author kept the behavior predictable by leaving those sockets out. And it's meant for prompt generation, not encoding: the MiniMax H3 conditioning encoder is truncated and unsuitable for text generation, which is exactly why you run this instead.