H3 Story Director — FL2VA Keyframe Storyboard

H3 Story Director - FL2VA Keyframe Storyboard is the most interesting of the Director variants, because it changes what you approve. Instead of planning only prompts, it plans ordered static keyframes and the video bridges between them - and each storyboard image is designed to do double duty: it's the opening image of its own video clip and the destination image of the clip before it. You approve a set of stills like a storyboard artist, and H3 (via FL2VA, first-to-last-frame animation) fills in the movement between them.

This is the "animatic before the animation" workflow, and for narrative video it's a genuinely strong way to work: you lock the beats visually, then let the video model handle the transition choreography rather than praying it invents a good shot sequence from prose alone.

How it works

It's the base Director plus a mandatory FL2VA keyframe rule set, and the rules are where the craft lives:

• Three vocabularies, cleanly separated. storyboard_prompt_prefix is still-only and assigns original <Picture N> references to S1–S3. prompt_prefix is video-only, never mentions any Picture tag, and refers to characters only as S1–S3. storyboard_prompt describes one static, high-quality composition per scene and may use the original Picture tags - but never movement, audio, captions, or multiple moments.
• The ordered storyboards are exact sequential keyframes. Storyboard N is the opening image of clip N and, when N > 1, also the destination image of clip N−1. Adjacent pairs are designed together, not independently.
• Before writing each non-final scene prompt, the node internally classifies the change to the next keyframe - continuous action, reframing/angle change, major action-state change, or location/time change - and expresses the right transition naturally inside the prompt. No extra JSON field; the classification is consumed and turned into prose.
• Continuous action preserves screen direction, body mechanics, momentum, camera trajectory, object contact and lighting while reaching the next pose. Reframing gets a motivated bridge: subject crossing the lens, foreground occlusion, whip pan, rack focus, orbit, push through a doorway. Major state changes get the full intermediate mechanics - weight transfer, hand and object paths, contact, reaction, the exact resulting pose. Location or time changes get a deliberate transition through darkness, a light wash, a doorway, a reflective surface, or match movement.

That's the whole mechanism: it turns "these two keyframes differ" into "here's how the camera and body get from one to the other," which is precisely the kind of temporal reasoning a text prompt usually gets wrong.

Inputs and outputs

Standard Director surface - api_key, model, story_idea, scene_count, scene_duration_seconds, draft_only, genre, motion_style, director_mode, video_sample_frames, four reference images, optional source_video - with the full nine outputs: plan_json, story_bible, synopsis, validation, usage_stats, credits_remaining, scene_prompt, mode_prompt, source_video_analysis.

Install

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

Restart, add an OpenRouter key, and remember the standard warning: the key field is masked but can live on in workflow metadata - scrub it before sharing.

The honest caveat: this is the most demanding Director variant to drive. It rewards you for actually reviewing the storyboards and adjusting scene count and duration before rendering. If you treat it as "write it and walk away," you're wasting the mechanism - the point is the stills you approve are the frames the video is contractually obligated to hit.