H3 Initial → Qwen

H3 Initial → Qwen is where a MiniMax H3 video starts. In I2VA mode, the whole job is "start the video from this image" - so this node builds the Qwen call that reads your Initial frame and turns it into the opening state the H3 prompt will describe. It's also used in Ref2VA as the weak composition reference: the picture that sets up framing and layout for shot one without fully dictating the characters.

Like the other "→ Qwen" nodes, it's a prompt builder, not an inference engine. It writes the system prompt, user prompt, and a deterministic config, and the ComfyUI_Simple_Qwen3-VL-gguf pack does the actual vision work. Nothing is called over the network, and there's no API key anywhere in this pack.

The one pass with four jobs

This is the busiest analyzer in the pack, because an Initial picture carries more than a Cast picture does. The prompt gives Qwen four jobs in a fixed order:

1. Find every clearly visible person or human-like character - ignoring people in posters, screens, reflections, and distant crowds.
2. List them in stable visual order (left to right, then top to bottom) as subject_0, subject_1, … with an appearance description for each.
3. Write one concise Japanese paragraph describing the visible opening state of the whole picture - pose, gaze, held objects, layout, lighting. Opening state, explicitly not future motion.
4. Write a short Japanese description of only the global rendering style (anime illustration, photoreal, 3D CG, …), or an empty string if there's no reliable style.

That split is the whole design: subjects, opening state, and style are separate fields on purpose, so you can edit one without clobbering the others. Up to eight people per picture, which is generous - and the prompt forbids inventing story, intention, or camera motion, because those are your job to write in the editor.

The interface

One optional input (base_config, for the model-selector override if you're hand-wiring), and the standard four outputs: system_prompt, user_prompt, config_override, and seed. The config pins temperature to 0.0 and turns thinking off, capped at 2048 tokens, JSON-only output. seed is always 0 - reproducible analysis is the point. Wire the outputs into Simple Qwen's node alongside your Initial image, exactly as the bundled workflows do.

Install and the honest gotchas

Install the whole pack through ComfyUI Manager (search "H3 Scribe") or git clone https://github.com/last-git/h3_scribe into ComfyUI/custom_nodes, then restart. You need ComfyUI_Simple_Qwen3-VL-gguf plus the Qwen3-VL GGUF and mmproj from the README (it recommends the 27B uncensored Q3_K_P build). The pack itself depends only on pydantic.

Where people trip: the description references people as local subject_0 names, and if Qwen references one you never listed, validation fails with "references unknown local subjects" - usually a sign the model got chatty, and worth just re-running. And remember this node is only for the Initial picture. If you're doing Cast-only Ref2VA, leave it unwired and tell Canonicalize References there's no Initial (initial_picture_number = 0); an Initial result wired into a cast-only graph is a layout error by design.