Ultimate H3 Cowboy Promptor

The H3 Auto Prompt Generator does one job brilliantly: character replacement from a reference image. The Ultimate H3 Cowboy Promptor is the node for everything else H3 can do. Where the other node is a focused tool, this one writes a MiniMax H3 prompt for any kind of shot - people, animals, objects, scenes, styles, actions, several subjects at once - and then hands the assets straight to the sampler. The README is clear that both are installed on purpose: one job done well, and one job done for the whole spec.

The promise is validated, not vibed: all six of MiniMax's own worked examples pass this node with zero errors, per the README.

The big idea: subjects are rows, not syntax

Instead of typing a monster prompt, you fill in rows. Each row is a kind (character, environment, animal, object, wardrobe, interface, effect, style, action, expression, pose), an optional name, and a description. character and environment sit at the top of the kind list and become the guide's own person and scene; the rest are the guide's other reusable kinds.

One rule runs the whole way through - row N is <Subject N> is subject_N_image is <Picture N>. So a skipped row leaves a hole rather than closing it up, and the node says so. A row needs no image (that's how you ask for a style or an action), and the typed subjects field behind the advanced panel is the escape hatch for a seventh subject or one that cites two pictures at once.

Two formats, one node

h3_mode picks the skeleton:

• ref - the six-section Ref2VA format from reference assets. This is the multi-subject mode.
• base_T2VA / base_I2VA / base_FL2VA / base_L2VA - the three-field base format: text alone, first frame, first+last frames, last frame. These need a different H3 checkpoint, which the h3_checkpoint_hint output names for you.

vlm_provider (same ten-option list as the Auto node) chooses the brain, target_description says what the target video should show, and cut_times wires Cut Detective in to pin the shot timeline.

The part that keeps everything honest: nothing is wired twice

The killer feature is the outputs. The images, video, and audio you plug in come back out as ref_image_1..6, ref_video (as IMAGE frames - what the sampler actually takes), ref_video_audio, and ref_audio, plus the width, height, and length that MiniMaxH3ReferenceToVideo wants and a label_map you can read against the prompt.

The point: the prompt's <Picture N> tags cannot drift from what the sampler receives, because the node hands you the exact assets it referenced. No "prompt says Picture 2 but I wired the wrong image" debugging session. length sits on H3's 17k+5 frame grid, and snap_duration_to_h3_grid (on by default) makes the prompt state the duration that grid really produces - ask for 2.00 seconds and H3 renders 2.33, so the prompt says 2.33.

Gotchas

• Backends are the same seven hosted/local options as the Auto node - this is a VLM-driven node, so it needs a provider (or a local model) to write the prose.
• Rows and images must stay in order. Slot 1 is <Picture 1>; if you wire subject images out of order, the prompt follows the slots, not your intent.
• snap_duration_to_h3_grid off means the prompt claims a shorter video than H3 produces. On by default, and there's a reason.

Install

One of ~69 nodes in TrentNodes:

# ComfyUI Manager: search "Trent Nodes"

# or:
cd ComfyUI/custom_nodes
git clone https://github.com/TrentHunter82/TrentNodes.git
cd TrentNodes && pip install -r requirements.txt

Pack install covers it; provider clients (anthropic/openai/google-genai) and keys are on you if you use hosted VLMs. If H3 character replacement is your whole world, the Auto node is leaner. The moment you want H3 to do more than swap a performer, this is the one you reach for.