H3 Story Director — Storyboard Cuts

H3 Story Director - Storyboard Cuts is the Director variant that refuses to let a video be planned as an afterthought. It plans independent shots joined by hard cuts, and for every scene it also produces one static storyboard frame - a locked first-frame composition that you can render as a still before you ever animate it. The generated sheet becomes the first-frame source for the FL2VA path. It's "design the shot, then set it in motion," which is a very different and often better way to control what H3 actually puts on screen.

The base Director writes moving-video prompts directly and hopes the model lands on a good first frame. This one splits the difference: the still is designed deliberately, approved visually, and then the video prompt is written to start from that exact approved image. If you've ever burned renders on a great prompt with a garbage opening frame, you'll feel this immediately.

How it works

Under the hood it's the base Director plus a block of mandatory storyboard rules appended to the system prompt - and those rules set up two completely separate reference vocabularies, which is the part that trips people up:

• storyboard_prompt_prefix - used only to generate still images. It assigns every connected original <Picture N> reference to stable subject labels S1, S2, S3, plus permanent visual rules.
• prompt_prefix - used only for video generation after the still exists. In it, <Picture 1> always means the generated storyboard image for the current scene, never an original uploaded reference. Original <Picture 2>, <Picture 3> and up are forbidden here entirely.

That role reversal is the core mechanism: the uploaded reference is who the subject is for the still; the generated still is what the video locks onto. The scene prompt then describes action, camera, dialogue and sound starting from that approved image, referring to characters only as S1–S3.

Every scene is an independent cinematic shot after a hard cut, so camera trajectory, composition and pose don't carry over - but identity, wardrobe, props, geography and narrative state do. And each storyboard_prompt is strictly one static composition: no temporal sequence, camera movement, audio, captions, or panel numbers.

The inputs and outputs

Same Director surface as the base node - api_key, model, story_idea, scene_count, scene_duration_seconds, steps, draft_only, genre, motion_style, director_mode, four reference images, optional source_video - and the same nine outputs, including plan_json, scene_prompt, mode_prompt, and story_bible.

The practical flow: render the storyboard stills, review them like an animator reviews a storyboard, then run the video pass with the approved still as the FL2VA first-frame source. That's where the "generated sheet becomes the FL2VA first-frame source" promise lands.

Install

Same pack, same command:

cd ComfyUI/custom_nodes
git clone https://github.com/elgalardi/ComfyUI-VisionPromptAssistant

Restart ComfyUI. OpenRouter key required, same key-in-workflow warning as the rest of the Director family - strip it before sharing.

Reach for this one when frame-level control matters more than speed: any project where the opening image of each shot is part of the idea, not a side effect. The extra step costs you a planning pass, but it buys a storyboard you can actually look at before you spend a single video render.