H3 Cast → Qwen

Despite the name, H3 Cast → Qwen doesn't call anything. It's a prompt builder: it writes the exact system prompt, user prompt, and sampling config that make Qwen analyze one Cast reference image the way MiniMax H3 needs it analyzed. The actual inference is done by the ComfyUI_Simple_Qwen3-VL-gguf pack, which this node feeds. No API key, no network, no "request" in the HTTP sense - just text that becomes the shape of a local Qwen call.

This is one of the two ANALYZE-stage builders in H3 Scribe (its sibling handles the Initial image). A Cast image in this pack means one clear person or human-like character - that's a hard contract, not a suggestion. The node's prompt tells Qwen to write a single appearance_ja description for exactly that one person and nothing else: no pose, no expression, no scene, no background, no lighting. Those go in other fields. The discipline matters because H3's prompt protocol separates subject appearance from scene and motion, and if Qwen smuggles "standing in a cafe" into a subject's appearance, that cafe gets baked into the character across every shot.

What comes out

Four outputs, all wired into Simple Qwen's node alongside the image:

• system_prompt and user_prompt - the instructions and the JSON schema Qwen must return.
• config_override - a JSON blob that pins sampling to deterministic settings: temperature 0.0, top_p 1.0, no thinking block, JSON-only output, capped at 2048 tokens. Same inputs, same analysis, every time. This is the small-and-obedient philosophy from the local-LLM playbook - you don't want a creative Qwen here, you want one that follows the format and stops.
• seed - always 0. Another determinism lock.

It takes one optional input, base_config, which is where you'd feed the JSON from the H3 Qwen Model Selector if you weren't using the bundled workflow (the selector's output merges into it). In practice the bundled graph handles all of that for you.

Install and the gotchas that actually bite

Install the whole pack via ComfyUI Manager (search "H3 Scribe") or git clone https://github.com/last-git/h3_scribe into ComfyUI/custom_nodes, then restart. The pack itself only needs pydantic; the heavy lifting is the ComfyUI_Simple_Qwen3-VL-gguf dependency plus a Qwen3-VL GGUF and its matching mmproj file.

Two real failure modes. First, the one-person contract: if your Cast image has multiple people, the prompt still describes only the single clearest person, and H3 Scribe does not do cross-image identity matching. A two-person cast photo is a bug, not a feature - crop to one person. Second, structured output: Qwen must return exactly one JSON object with no thinking block and no code fences (a single fence is tolerated). If it leaks a <think> block or prose, the parser throws a clear "Qwen returned non-JSON text" style error. The recommended uncensored Qwen3-VL GGUF from the README behaves; a model that refuses to comply with the JSON schema will make you chase your tail. That's the Qwen model's fault, not this node's.