Video Prompt Pixaroma

Video Prompt Pixaroma is the "type it in plain English, get a finished MiniMax H3 prompt" node. H3 wants its prompts in a particular shape - named sections, a soundscape, music, and a strict way of writing anything a person says out loud - and getting that shape wrong quietly spoils the clip. This node does all of it from a sentence like "a blacksmith hammers glowing steel in a dark forge", on your own machine, through a vision language model in your text_encoders folder. No account, nothing sent anywhere.

What makes it more than a prompt formatter is the frames output. H3 accepts durations on a specific frame pattern, and the node hands back a frame count already snapped to it - so the video is exactly as long as the prompt was written for. Getting those two out of step is the easiest way to waste a render, and this kills the whole class of bug.

You don't pick a mode - the wires do

• Wire nothing and it writes text-to-video.
• Wire one picture and it looks at it and animates it (first-frame mode).
• Wire two pictures and it writes the journey between them, joining them for you so they can't end up the wrong way round (first-and-last mode).

Inputs

• first_frame (IMAGE, optional) - the picture the video starts on; connecting it switches to first-frame mode.
• last_frame (IMAGE, optional) - the picture the video ends on; connecting this as well switches to first-and-last mode.
• clip (optional) - wire a CLIP Loader here to use that model instead of the one chosen in settings. Handy for sharing a single loaded model between several of these nodes.

Outputs

• text - the finished H3 prompt; wire it into the prompt/text input of your H3 node.
• frames (INT) - how many frames to render, already adjusted to H3's pattern; wire it into the length input.
• seconds (FLOAT) - the real duration (frames ÷ frame rate), for anything that has to line up with the video, like the length of an audio track.

Everything the node says is editable: the wording it follows lives in the settings, kept outside the plugin folder so updates never overwrite your edits, and you can switch the length instructions off entirely and use your own for any video model you like. A Free VRAM switch hands memory back to your video model once the prompt is written.

The model it needs

Because it has to see pictures, it needs a vision model: the reference is qwen3-vl-8b-heretic-1.3.0_fp8_e4m3fn.safetensors (about 10 GB, 12 GB+ cards) with a 4B build for 8 GB. Drop it in ComfyUI/models/text_encoders/; the node finds a suitable model by itself if you have one.

Install

Pixaroma pack, standard drill: ComfyUI Manager → search Pixaroma → install → restart, or:

cd ComfyUI/custom_nodes
git clone https://gitlab.com/pixaroma/comfyui-pixaroma.git

The honest note: this node writes the prompt; it doesn't run H3. Pair it with your H3 sampler workflow, and give the vision model and the video model separate VRAM headroom - running a 10 GB vision encoder and H3 on the same card is where memory errors come from. And after an update, hard-refresh the tab (Ctrl+Shift+R / Cmd+Shift+R) if a node looks stale.