APNext H3 Crossover Writer

Every video model can put two characters in a room. Putting Caitlyn from Arcane in a room with Bruce Banner - and keeping them them - is where naive prompting dies, which is why APNext H3 Crossover Writer exists. It takes a cast (from H3 Characters nodes, or typed by hand) plus your creative brief, and has the local Claude Code CLI write 1–10 crossover scenes, each a complete four-section T2VA prompt (subject_definitions / integrated_multimodal_description / overall_soundscape / non_diegetic_music).

The rules it writes under aren't vibes. They're distilled from actually-rendered crossover productions and shipped in data/h3/guide_crossover_en.md: actor pinning, <Subject N> speaker binding, silence mandates, not in frame isolation, positioned two-shots, no dead air, grounded entrances, hand-offs between scenes. No title cards unless you ask. That's the part you can't get from just pasting "crossover" into a prompt box - the guide is the system prompt, so every scene comes out of the model already knowing who's where and who gets to talk.

The inputs that matter

• direction - your creative brief: premise, tone, where it happens, what must happen, running gags, who should clash. This is the actual writing; everything else is scaffolding.
• cast_1..4 - cast outputs from H3 Characters nodes. Chain several Characters nodes through cast_in to stack a big cast onto one socket, or use several sockets. extra_cast takes hand-typed lines (Character (played by Actor) from Show) merged in.
• scene_count - 1–10.
• duration_mode / scene_duration - fixed length per scene, or let Claude pace each scene 5–15s.
• continuity_mode - Independent clips (each scene its own T2V clip, hard cuts) vs Continuous chain (written for C2V / motion-context chaining, scene N+1 opening on scene N's last frame). This decides whether you stitch or chain downstream.
• wardrobe / locations - the locks: one line per character/place, copied word-for-word into every shot the character is on screen in, every scene set in that place. Empty = the model fixes its own in the synopsis and repeats them. enforce_wardrobe (on by default) parses the result and fires one repair turn in the same session if any shot dropped an anchor - the result shows up in info as wardrobe: ok (2 locked) or repaired 6 -> 0.
• image_1..9 - reference pictures. Downscaled copies go to Claude so it can recognise who/what each picture is (use image_notes - Image 1: Sheldon); pictured characters get bound to their <Picture k> and their wardrobe lock is taken from the picture. The originals pass through to the matching outputs, ready for MiniMax H3 Reference to Video.

Outputs: scenes (list) and durations (list) - one per scene - plus scenes_text, synopsis, the merged cast, scene_count, session_id, info, and the image_1..9 passthrough.

Rendering options

A video node fed scenes runs once per element: one queue, every scene. For iteration, H3 Scene Pick grabs one scene by index. For one output file, H3 Scenes Join. For a seamless single take, the pack's h3_crossover_contex_chain.json runs the writer in continuous-chain mode into H3 Scenes → Contex Loop Plan. The h3_crossover_batch.json, h3_crossover_pick_one.json and h3_llm_backend_crossover.json examples cover the rest.

Installing

Part of dagthomas/comfyui_dagthomas ("SDXL Auto Prompter"): ComfyUI Manager → comfyui_dagthomas, or clone into custom_nodes + pip install -r requirements.txt, restart. Node under APNext → H3.

Requirements are Claude Code installed and logged in (subscription seat by default via use_subscription; flip it to bill an API key instead if you'd rather). Multi-scene runs are the slow end - several scenes × 25–60s each plus the enforcement pass. If you'd rather run free, an H3 LLM Backend on the llm socket moves the whole thing to Ollama, with research ignored but locks, sessions and outputs otherwise identical.