Image → Scene Prompt

Convert a static ink painting into a motion prompt for image-to-video generation. The motion prompt tells the video model what CHANGES between frame 1 and the last frame — not what the image looks like.

Required Input

The original image prompt from verse-to-prompt is mandatory. This is the source of truth for what exists in the painting.

Acceptable inputs:

• The full verse-to-prompt output (preferred — has style, prompt, and translation)
• The image prompt string alone (e.g., "A river god rises from churning rapids...")
• A hexagram key (e.g., 44-44) — you look up the prompt from data/yilin/prompts/
• A Records card number — in 8bitoracle-brand, find the card's plate in records/vol1/cardlist.json, then the row in records/art/prompts-manifest.json whose output basename matches it

Never generate a motion prompt from a verbal description, screenshot, or memory of a painting. If you don't have the original image prompt, stop and ask for it.

Retouched plates log the wrong prompt. An inpaint composite (…-fix.png, model …/fill) logs only its fill prompt — "bare plaster wall", "plain hazy sky" — not the scene. Use the prompt of the base render it was composited onto (the same name without -fix). Then look at the final plate once: retouching can remove elements the base prompt names.

The Correspondence Rule

Every element named in the motion prompt MUST exist in the original image prompt. Read the image prompt, identify its movable elements (water, wind, figures, light, particles), and build motion ONLY from those.

# Original image prompt (source of truth):
"A river god rises from churning rapids, mouth open in a great shout,
arms spread to bar the crossing. Foreground, waves crash against a stone
ferry landing, spray flying diagonally. Midground, a traveler reins in
his horse at the water's edge, startled."

# Movable elements: rapids, waves, spray, horse, storm clouds
# NOT movable: river god (static pose), stone landing (architecture)

# Good motion prompt — only references elements from the image:
"Water surges forward, spray rising. Horse steps back nervously.
Storm light breaks through clouds. Static camera."

# Bad — adds elements not in the image:
"Birds scatter from the riverbank as fish leap from the water."

Your Task

Given the original image prompt, produce:

1. A motion classification with a one-line rationale
2. A movable elements list extracted from the image prompt
3. A 15-25 word motion prompt — motion and change only, no scene description
4. A duration recommendation (5s or 10s) with cost estimate
5. Optional: loop instruction if the clip should seamlessly loop

The 5 Motion Categories

The 7 Motion Rules (MANDATORY)

1. Only describe what changes — The image already shows the scene. Your prompt is NOT a scene description. It is a set of technical instructions for what moves and how. If the image shows it, do not write it. If the image cannot show it (motion, time passing, light shifting), that is your prompt.
2. Derive from the image prompt — Read the original image prompt. List every element that could plausibly move (water, wind, fabric, animals, smoke, light). Pick 1-2 as your motion sources. Never invent elements.
3. One primary motion, one secondary — Video models fragment with competing directives. Pick the dominant movement and one ambient layer. Never three simultaneous motions.
4. 15-25 words, literal structure — A motion prompt is a technical instruction set, not poetry or prose. Subject → Action → Camera → Settling cue. Every word must earn its place.
5. Speed is always slow — Ink paintings don't move fast. Use "gradually", "gently", "slowly". Even kinetic-burst builds slowly then releases one quick moment.
6. Camera is usually static — For i2v, static camera produces the most stable results. Only use camera motion (slow push-in, gentle pull-back) when the painting has strong depth corridors. Never orbit or pan.
7. End with a settling cue — The last few words prevent abrupt endings: "settling into stillness", "light fading", "motion easing". For loops, replace with: "End frame matches start frame for a seamless loop."

Prompt Structure

[Camera], [primary motion]. [Secondary motion]. [Settling cue].

Keep it this tight. The model performs best with literal, specific instructions.

Anti-Patterns

Duration & Cost Guidance

Platforms

With a QwenCloud Token Plan, use it: about $0.38 a clip if the weekly credits are used (see QwenCloud Token Plan). Without one, or when the plan's weekly quota runs out, pay per clip:

fal prices checked 2026-09-14. At 720p fal and Replicate cost the same; fal runs HappyHorse 1.1, and its 1080p costs less.

Default to Happy Horse, 720p, 5s ($0.70/clip pay-per-clip).

Budget planning (720p, 5s):

API Input Reference

import replicate

output = replicate.run(
    "alibaba/happyhorse-1.0",
    input={
        "prompt": "Water surges forward, spray rising. Horse steps back. Static camera.",
        "image": "https://cdn.example.com/44-44.webp",
        "duration": 5,
        "resolution": "720p",
        "seed": 42,
    }
)

Happy Horse parameters:

• image — first-frame image (jpg/png/bmp/webp, ≤10MB, each side ≥300px)
• duration — 3-15s integer
• resolution — 720p or 1080p
• seed — 0-2,147,483,647

Seedance 2.0 additional parameters:

• last_frame_image — optional end frame (controlled transitions)
• generate_audio — native audio sync
• reference_images — up to 9 style/character references (cannot use with image)
• duration supports -1 for intelligent duration

QwenCloud Token Plan (HappyHorse 1.1)

Use this path when the account has a QwenCloud Token Plan (keys start with sk-sp-) instead of Replicate. The plan costs $68 a month and gives 40,000 credits a week for four weeks. It spends credits, not dollars per clip.

The runner code is in 8bitoracle-next, not sixlines-ios. Run labels say sixlines-ios-… because that app shows the videos. Nothing in sixlines-ios calls Qwen.

Credentials. The key and host are in ~/.claude/settings.json.qwen. Export them for the runner:

export QWEN_TOKEN_PLAN_API_KEY=$(jq -r .env.ANTHROPIC_AUTH_TOKEN ~/.claude/settings.json.qwen)
export QWEN_TOKEN_PLAN_BASE_URL=$(jq -r .env.ANTHROPIC_BASE_URL ~/.claude/settings.json.qwen)
pnpm exec tsx scripts/run-qwen-yilin-video-smoke.ts path/to/input.json

Request. {origin} is the origin of that base URL, https://token-plan.ap-southeast-1.maas.aliyuncs.com.

POST {origin}/api/v1/services/aigc/video-generation/video-synthesis
Authorization: Bearer sk-sp-…
X-DashScope-Async: enable
{ "model": "happyhorse-1.1-i2v",
  "input": { "prompt": "<motion prompt>",
             "media": [{ "type": "first_frame", "url": "<image url>" }] },
  "parameters": { "duration": 5, "resolution": "720P", "watermark": false, "seed": 2252 } }

GET {origin}/api/v1/tasks/{task_id}    ← no async header; repeat every 5 s until SUCCEEDED or FAILED

Resolution is uppercase (720P); Replicate uses lowercase. For text-to-video, use happyhorse-1.1-t2v and omit media. The result URL is in output.video_url. Download it at once to durable storage; do not store the provider URL as the link.

Audio (HappyHorse 1.1)

Every clip comes with sound. HappyHorse makes the audio and the video in one pass, and no parameter turns the audio off. Records Vol I clips carry a stereo AAC track at 24 kHz.

If the prompt names no sound, the model invents it from the picture and the motion. Records card 23 asked only for "the chancellor pleads… the innkeeper shakes his head", and the clip came back with Mandarin dialogue on that subject. That can be good. It can also put words in a character's mouth that the story does not support. So decide the sound, and write it.

• Lip-sync languages: English, Mandarin, Cantonese, Japanese, Korean, German and French. Classical Chinese is Mandarin read aloud; say "in Mandarin".
• Keep each line short. A 5 s clip holds about one short sentence. A face turned to the camera with the mouth visible gives the cleanest lip-sync. Most Records plates show figures small or from the side, so on those cards prefer ambience or No dialogue.
• The 15–25 word rule is for the motion. The audio clause is extra. Put it after the motion as its own sentence, starting Audio:, so it does not compete with the motion instructions.
• Viewers hear it only when unmuted. Browsers autoplay only muted video. Review the sound on purpose; a muted review misses it.

"Vendor guides" means reseller guides for HappyHorse 1.1 (Morphic, PixVerse, SeaArt), not yet tested on the Token Plan. After the first tested clip, change that row to what it showed.

Check the speech with Whisper. You cannot hear a clip, but Whisper can transcribe it:

python3 -c 'from faster_whisper import WhisperModel as W; s,_=W("small",device="cpu",compute_type="int8").transcribe("clip.mp4",language="zh"); [print(f"{x.start:.1f}-{x.end:.1f} {x.text}") for x in s]'

• Whisper writes homophones for classical Chinese: 丞相誤邪 came back as 成像物協. Compare the sound, not the characters.
• On a clip with no words, Whisper makes up a subtitle credit, such as 「字幕by索兰娅」. That means no speech, not a bad take.

Check song, music and effects with a spectrogram. Whisper misses sound that has no words: the Chu song on Records card 47 came back as a made-up credit.

ffmpeg -i clip.mp4 -lavfi "showspectrumpic=s=900x300:legend=0:fscale=lin:stop=4000" spec.png

Read the picture. Singing and pitched instruments show as stacked horizontal bands that step in pitch. Drums and crackle show as vertical or broadband streaks. Ambience with nothing else is one low band. Compare the new take with the old one.

A sound with no words can be the whole story (a song, a rooster, a drum). Write it in the Audio: clause after No dialogue. Card 47's Many distant voices sing a slow, mournful Chu folk song from every side produced the song. The first take, with no audio clause, had none.

Gotchas

Cost (Personal Pro: $68 a month for 40,000 credits a week over four weeks, 160,000 in all; clip cost measured 2026-09-12): one 5 s 720P i2v clip costs about 900 credits (2.25% of a week), so a week buys about 44 clips and a month about 177 — about $0.38 a clip, against $0.70 on fal. That price holds only if the credits are used: $68 buys about 97 clips on fal, so the plan is cheaper only above about 24 clips a week. The dashboard shows one decimal place, so one clip reads as 2.2 or 2.3 points; measure across several clips. A clip takes about 100 s. The runner does not check the balance: before a batch, divide the remaining credits by 900.

Output Format

Original image prompt:

A river god rises from churning rapids, mouth open in a great shout, arms spread to bar the crossing. Foreground, waves crash against a stone ferry landing, spray flying diagonally. Midground, a traveler reins in his horse at the water's edge, startled. Beyond, the far bank is barely visible through mist and foam. Amber light breaks through storm clouds above the raging torrent. Chinese ink painting.

Motion: kinetic-burst — churning rapids and rearing horse provide two strong motion anchors

Movable elements: rapids/waves, spray, horse, storm clouds, amber light

Prompt (5s):

Water surges forward, spray rising diagonally. Horse steps back. Storm light breaks through clouds. Static camera.

Duration: 5s — single tension moment, no arc needed
Cost: ~$0.70 (Happy Horse 720p)
Loop: No — the scene has directional energy, not cyclical

Reference Prompts

slow-reveal

From: "A frozen river valley under a dark sky split by driving snow..."
Movable: snow, willow branches, lantern light

Slow push-in. Snow thickens. Willow branches tremble. Lantern light pulses warmer. Settling into stillness.

living-landscape

From: "An empty market square in a river town, rain falling in gray sheets..."
Movable: rain, puddle reflections, paper lantern

Rain falls steadily. Puddle reflections ripple. Red lantern sways. Light dims toward dusk. Static camera.

figure-breath

From: "A betrothal ceremony in a courtyard dusted with autumn frost..."
Movable: incense smoke, red silk, lantern light, frost glitter

Incense smoke curls upward. Red silk stirs in faint breeze. Lantern light shifts slowly. Static camera.

kinetic-burst

From: "A massive tiger crouches low at a rocky stream..."
Movable: tiger muscles, stream water, spray

Tiger shifts weight forward. Stream surges against rocks, spray catching light. Static camera. Tension building.

celestial-drift

From: "An immortal sage ascending from a dark mountain pool on the back of a celestial horse..."
Movable: horse's mane, cloud wisps, starfield reflection, gold light

Slow crane up. Horse's mane streams upward. Gold stars pulse brighter. Cloud wisps trail behind.

looping example

From: "A gnarled pine leans from a cliff face..."
Movable: mist, river light, vermillion leaf

Mist drifts through gorge. River light ripples. Leaf trembles. End frame matches start frame for a seamless loop. Static camera.