RunningHub MiniMax H3 T2VA Text Encode (Legacy)

Every video model has a moment where your words become numbers. For MiniMax H3, that moment is RHMiniMaxH3T2VATextEncode: it takes the loaded Qwen3-VL text encoder and a prompt, and produces the conditioning that steers the video-plus-audio generation. It's a two-input, one-output node that was workhorse of the old T2VA chain - and it's now legacy, because the modern generation node does the same job inside itself.

What it does

Two required inputs, nothing else:

• h3_text_encoder - the handle from RHMiniMaxH3TextEncoderLoader (or the legacy direct loader).
• prompt - multiline text. Empty prompts are rejected outright; this isn't a node where blank means "no conditioning," it throws.

One output: conditioning (type MINIMAX_H3_CONDITIONING), which wires into the encode node in the old chain. The encoding itself runs through Qwen3-VL 32B - the same vision-language model that reads reference media - which is why H3's prompts can mix direction for the picture and the sound in one block. It's a heavy encode, and it's the same encode the modern generation nodes cache so re-runs don't pay for it twice.

Why it's deprecated

RHMiniMaxH3VideoGen calls this exact logic internally: give it a prompt and it loads the text encoder, encodes, and caches the conditioning - you never see this node. The pack's design is explicit about it: the legacy nodes are registered purely so old serialized workflows still instantiate and execute, but they're hidden from search and the node tree. If you're building anything new, don't go hunting for this node; just use RHMiniMaxH3VideoGen or RHMiniMaxH3RefGen and type your prompt into the widget.

When it's actually useful

Maintaining an old workflow is the main reason. If you're doing that, be aware the prompt text here should follow MiniMax's own writing guide (the official docs are bundled in the pack's docs/prompt_guide.md): write the body in English, keep <d>-quoted dialogue and visible text in their original language, and treat the picture and audio as one timeline. Also note the encode is dynamicPrompts-enabled, so wildcard-style prompt expansion works if you use a supporting frontend.

Install

Standard pack routine:

cd ComfyUI/custom_nodes
git clone https://github.com/RH-RunningHub/ComfyUI-RH-MiniMax-H3.git
pip install -r ComfyUI-RH-MiniMax-H3/requirements.txt

Restart, then the ~95 GiB INT8 ConvRot bundle into models/MiniMax-H3-INT8-CONVROT/ (the Qwen3-VL weights live inside it). If you hit a "prompt cannot be empty" error you didn't expect, it's the widget wired to a string that resolved to whitespace - the check is on stripped text.