# H3 Keyframe Film

> Write runnable MiniMax H3 keyframe-film prompts — multi-keyframe montages, first-last frame transitions (FL2VA), final-frame landings (L2VA), and I2VA first-frame animation. Use when the user has reference images (keyframes, character art, product shots, final poses) and wants the image identity locked while text drives the motion. Follows base-en.txt exactly, including the keyframe-alignment instruction lines. Pure-text requests are redirected to h3-promo-film; reference-video motion goes to reference-motion-transfer.

- Skill: `gordonlu/h3-keyframe-film` (Agent Skill, multi-file: 6 files)
- Install (CLI): `npx skillmds@latest add gordonlu/h3-keyframe-film`
- Raw SKILL.md: https://api.skillmd.com/api/skills/gordonlu/h3-keyframe-film/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: gordonlu (https://skillmd.com/u/gordonlu)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/gordonlu/h3-keyframe-film

---


# H3 Keyframe Film Studio

Use this Skill when the user has **images but no video** and wants a storyboard-style short from MiniMax H3: "cut through these four keyframes", "transition between these two poses", "land exactly on this final frame". Images lock the identity; text locks the motion.

Core principle: **the alignment line decides what the video is anchored to; the timecoded beats decide what happens between anchors.**

Prompts follow `h3-prompt-writing` (`base-en.txt`). The alignment instruction lines are fixed templates — copy them verbatim.

## 1. Trigger

Use when: I2VA first-frame requests; FL2VA pose-to-pose transitions; L2VA "end on this exact frame" requests; multi-keyframe montages (up to 4 frames); image + text motion description in any language.

Do **not** use when: no images at all (→ `h3-promo-film`), a reference *video* is the motion source (→ `reference-motion-transfer`), the user wants to edit an existing clip, or the user has **exactly two images and nothing else** (→ `h3-fl2va-film`, the FL2VA-only studio).

## 2. Intake

Confirm: ① the image(s) and their role (start frame / end frame / keyframe sequence), ② aspect ratio & duration (4–15s), ③ motion description with beats, ④ style & scene (comes from the image + style sentence), ⑤ on-screen text, ⑥ audio policy.

## 3. Goal

One approved video where the subject's identity matches the reference images in every frame and the motion follows the written beats exactly, ending per the mode contract (I2VA: first frame locked; FL2VA: both frames locked; L2VA: final frame locked).

## 4. Mode Selection

| Assets | Mode | Alignment line |
| --- | --- | --- |
| First-frame image + text motion | **I2VA** | `For the target video, at 0.00 seconds into the target video, <Picture 1> (from [Shot 1]) is fully referenced.` |
| Start + end pose images | **FL2VA** | `How the reference pictures align with the target video — Picture 1 (from Shot 1) aligns with the 0.00-second mark ...; Picture 2 (from Shot N) aligns with the S.SS-second mark ...` |
| End-pose image only | **L2VA** | `How the reference pictures align with the target video — <Picture 1> (from [Shot N]) aligns with the S.SS-second mark ...` |
| Up to 4 keyframes, sequential | **I2VA multi-frame** | One `fully referenced` clause per frame, each pinned to its timestamp |

## 5. Prompt Construction

After the alignment line and a blank line, three fields:

1. `integrated_multimodal_description` — style sentence → `[Shot N]` (no timestamp) → timecoded beats `From X.XX to Y.YY seconds, ...`. Identity anchors (3–5 per subject: silhouette, colors, signature props) must recur across beats. Motion path structure: initial state → middle changes → gradual approach → final state landing on the end frame.
2. `overall_soundscape` — per-beat diegetic sound.
3. `non_diegetic_music` — genre, BPM, curve; `N/A` when unused.

Rules: end timestamp must equal the total duration with two decimals (`S.SS`); single-shot preference for FL2VA / L2VA; dialogue/on-screen text keep original language; 350–500 English words.

Ready-to-run examples: [`references/prompt-library.md`](references/prompt-library.md).

## 6. Generation & Review

- Model: **MiniMax H3**, ratio & duration as planned; 2K only after approval.
- Checklist: identity holds in every frame (no feature leak from other reference images) / each beat appears in order with correct timing / end frame matches the locked picture / camera matches the written move / audio matches policy. Score pass / partial (name the beat) / fail.

## 7. Retry Strategy

| Symptom | Fix |
| --- | --- |
| Identity drifts mid-video | Add 1–2 identity anchors; re-state the image as a mid-video keyframe if I2VA multi-frame |
| End frame not landed | Re-check the alignment line timestamp; extend the final approach beat |
| Motion too generic | Add timecoded micro-beats; shorten the windows |
| Wrong subject element leaks in | Remove the leaked attribute from every beat; restate `fully referenced` clauses |
| Camera inconsistent | State the move inside each beat sentence |

Escalate after two failed retries of the same layer — offer fewer keyframes or a simpler motion path.

## 8. Output

Deliver: final video file; the prompt with alignment line exactly as used; settings record (mode, ratio, duration, audio policy); review notes.

## Boundaries

Image-to-video only. No images → `h3-promo-film`. Reference-video motion → `reference-motion-transfer`. Do not animate real identifiable persons without consent, brand logos, or copyrighted characters; abstract or decline.