Abliteration Vision Prompt

If Vision Prompt Assistant is the node that runs a local vision model to write your MiniMax H3 prompt, Abliteration Vision Prompt is its cloud sibling: same job, same system/user prompt split, same three-image reference mapping - but the generation happens on Abliteration.ai's OpenAI-compatible vision endpoint, not your GPU. The pitch is simple: the hosted abliterated-model uses zero of your VRAM, which matters when MiniMax H3 itself is about to eat every byte of it.

This is a text node, not a video node - it returns a STRING prompt, no media. You're using it exactly where you'd use the local version: reference-aware H3 prompting without the encoder sitting in memory.

How it works

You paste an API key from Abliteration.ai, type a user and system prompt, and optionally connect up to three images. Images are mapped in connection order to <Picture 1>, <Picture 2>, <Picture 3>, resized to image_max_dimension (1024 by default - a decent balance of detail, latency, and token cost), and JPEG-encoded before upload.

One genuinely thoughtful detail: when you connect two or three images, the node stitches them into a single labeled contact sheet before uploading, and tells the model to treat each labeled panel as a separate reference. That preserves the MiniMax <Picture n> mapping even on API backends that only process the first visual block in a request.

The inputs that matter

• api_key - required, masked in the UI.
• user_prompt / system_prompt - same roles as the local node; the defaults are tuned for H3 Reference to Video.
• image_0 / image_1 / image_2 - up to three references.
• thinking - leave false for faster prompt enhancement; it's a reasoning toggle, not a quality switch.
• max_tokens - 256 default, up to 4096.
• image_max_dimension - lower it to shave token cost on a big reference stack.

Outputs

• generated_text - the prompt. Wire this to your H3 conditioning or to Preview Vision Prompt.
• usage_stats - a human-readable string with input/output/total token counts from the service.
• credits_remaining - credit balance after the request, credits used, and estimated USD cost when the API reports them. Handy if you're watching spend.

Install

Same pack as everything else here: ComfyUI Manager → search Vision Prompt Assistant, or:

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

Then restart. No model downloads - that's the whole point.

The honest warning

The API-key field is visually masked, but - like most API nodes in ComfyUI - the value can still end up serialized inside a saved workflow. Before you share a workflow JSON or an image/video whose metadata embeds it, strip the key out. If it ever leaks, rotate it. Same advice that applies to every key-carrying node in the graph; the difference here is that this key is a paid meter.

Set thinking off, keep image_max_dimension at 1024, and this is a fast, cheap way to get reference-aware H3 prompts without touching your VRAM budget. If you'd rather keep everything local, use Vision Prompt Assistant instead - you already have it in the same pack.