renoise-gen（雷诺斯生成器）

Renoise CLI — Tool Reference

CLI for the Renoise AI video/image generation platform. This skill covers how to use the tools. For creative decisions (what to generate, how to write prompts, anchoring strategy), see the director skill.

Platform URL: https://www.renoise.ai — NEVER renoise.com.

Quick Start

# Text-to-Video — 15s finished cut
node ${CLAUDE_SKILL_DIR}/renoise-cli.mjs task generate \
  --prompt "[0-5s] Close-up of a cat on the moon, slow push in. [5-12s] The cat dances under twinkling stars, smooth orbit. [12-15s] Wide pull back revealing the full lunar landscape, frame holds steady." \
  --duration 15 --ratio 16:9

# Image-to-Video — upload a reference image, then generate
node ${CLAUDE_SKILL_DIR}/renoise-cli.mjs material upload /path/to/photo.jpg
node ${CLAUDE_SKILL_DIR}/renoise-cli.mjs task generate \
  --prompt "The product rotates slowly on a white pedestal, soft studio lighting, cinematic." \
  --materials "42:ref_image" --duration 10 --ratio 16:9

# Generate Image
node ${CLAUDE_SKILL_DIR}/renoise-cli.mjs task generate \
  --prompt "A cute cat sitting on a crescent moon, watercolor style, dreamy atmosphere" \
  --model nano-banana-2 --resolution 2k --ratio 1:1

# Generate Image with Text/Typography (use gpt-image-2 for best prompt-following on text/logos)
node ${CLAUDE_SKILL_DIR}/renoise-cli.mjs task generate \
  --prompt "Hero product poster with bold headline 'MIDNIGHT BLOOM' in serif type, centered, minimalist layout" \
  --model gpt-image-2 --resolution 2k --ratio 16:9

Supported Models

Choosing an image model:

• nano-banana-2 — default, cheapest, flexible ratios (incl. extreme 8:1 / 1:8 banners).
• nano-banana-pro — higher fidelity; use for hero / final frames.
• midjourney-v7 — strongest stylization; pass --ratio only (no resolution).
• gpt-image-2 — best prompt-following for text/logos/typography in image.

CLI Commands

node ${CLAUDE_SKILL_DIR}/renoise-cli.mjs <domain> <action> [options]

Domains: task, material, asset, character, credit

Credit

node ${CLAUDE_SKILL_DIR}/renoise-cli.mjs credit me                    # User info + balance
node ${CLAUDE_SKILL_DIR}/renoise-cli.mjs credit estimate --duration 15 # Estimate cost
node ${CLAUDE_SKILL_DIR}/renoise-cli.mjs credit history                # Transaction history

Task — Generate & Manage

# Generate (create + wait + return result)
node ${CLAUDE_SKILL_DIR}/renoise-cli.mjs task generate \
  --prompt "..." --duration 15 --ratio 16:9 \
  [--materials "..."] [--characters "..."] [--tags "..."]

# Create only (returns immediately)
node ${CLAUDE_SKILL_DIR}/renoise-cli.mjs task create \
  --prompt "..." --duration 15 --ratio 16:9

# Management
node ${CLAUDE_SKILL_DIR}/renoise-cli.mjs task list [--status completed] [--tag project-x]
node ${CLAUDE_SKILL_DIR}/renoise-cli.mjs task get <id>
node ${CLAUDE_SKILL_DIR}/renoise-cli.mjs task result <id>
node ${CLAUDE_SKILL_DIR}/renoise-cli.mjs task wait <id> [--interval 10] [--timeout 600]
node ${CLAUDE_SKILL_DIR}/renoise-cli.mjs task cancel <id>              # Pending only

# Chaining — download result → upload as material (one step)
node ${CLAUDE_SKILL_DIR}/renoise-cli.mjs task chain <id>
# → prints new material ID, ready for --materials "ID:ref_video"

# Tags
node ${CLAUDE_SKILL_DIR}/renoise-cli.mjs task tags
node ${CLAUDE_SKILL_DIR}/renoise-cli.mjs task tag <id> --tags a,b,c

Parameters for generate / create:

Task statuses: pending → assigning → assigned → queued → running → completed / failed

Material — Upload & Browse

node ${CLAUDE_SKILL_DIR}/renoise-cli.mjs material upload /path/to/file.jpg
node ${CLAUDE_SKILL_DIR}/renoise-cli.mjs material upload /path/to/clip.mp4 --type video
node ${CLAUDE_SKILL_DIR}/renoise-cli.mjs material list [--type image] [--search cat]

Asset — Register for Face/Character Use

Register uploaded images as Ark assets so they bypass privacy detection when used as reference_image.

# One-step: create + wait until active (~30-60s)
node ${CLAUDE_SKILL_DIR}/renoise-cli.mjs asset register <material_id> --name "Character Name"

# Step by step
node ${CLAUDE_SKILL_DIR}/renoise-cli.mjs asset create <material_id> --name "Character Name"
node ${CLAUDE_SKILL_DIR}/renoise-cli.mjs asset wait <id>

# Manage
node ${CLAUDE_SKILL_DIR}/renoise-cli.mjs asset list
node ${CLAUDE_SKILL_DIR}/renoise-cli.mjs asset get <id>
node ${CLAUDE_SKILL_DIR}/renoise-cli.mjs asset delete <id>

Asset statuses: pending → processing → active / failed

Character — Browse Platform Library

node ${CLAUDE_SKILL_DIR}/renoise-cli.mjs character list [--category female] [--search Jasmine]
node ${CLAUDE_SKILL_DIR}/renoise-cli.mjs character get <id>

Material Roles

Three mutually exclusive modes for visual input. Do NOT mix modes.

Mode-level exclusion: first_frame cannot be mixed with ref_image. Within multimodal reference mode, ref_image + ref_video + asset:ID:reference_image combine freely — they use different API fields (id vs user_asset_id).

Available roles in multimodal reference mode

Combine as needed per segment:

# Character + continuity + environment
--materials "asset:27:reference_image,42:ref_video,99:ref_image"

# Character only
--materials "asset:27:reference_image"

# Environment only (B-roll)
--materials "99:ref_image"

API representation:

"materials": [
  { "user_asset_id": 27, "role": "reference_image" },
  { "id": 42, "role": "ref_video" },
  { "id": 99, "role": "ref_image" }
]

Image requirements (for first/last frame and ref_image)

• Format: jpeg, png, webp, bmp, tiff, gif
• Aspect ratio (W/H): 0.4 – 2.5
• Dimensions: 300 – 6000 px per side
• Size: < 30 MB

Privacy detection

Images with human faces passed as ref_image will be blocked with PrivacyInformation error. Two ways to bypass:

• User Asset: asset register → asset:ID:reference_image
• Character Library: --characters "ID"

Never pass raw face images as ref_image.

Material Pool (Batch Ingest)

Scan a folder, upload, analyze with Gemini, output structured material-pool.json:

node ${CLAUDE_SKILL_DIR}/scripts/material-ingest.mjs ./materials/

Auto-match materials to shots:

node ${CLAUDE_SKILL_DIR}/scripts/match-materials.mjs --pool material-pool.json --shots project.json

Face detection: materials with has_face: true are automatically excluded from ref_image matching.

Prompt Basics

The renoise-2.0 model responds best to:

• English only — non-English text or tag lists degrade quality
• Narrative paragraphs — complete descriptive sentences, not comma-separated keywords
• Specific over abstract — "a golden retriever running through shallow ocean waves at sunset" beats "a dog on a beach"
• Positive language — describe what you want, not what you don't want

Structure: Subject (appearance) + Action (what happens) + Camera (movement) + Scene (environment) + Style (visual style)

Shot density for time-annotated prompts:

• 5s: 1 shot, single action + camera
• 10s: 2–3 shots
• 15s: 3–4 shots
• End the last segment with "frame holds steady" for clean endings

For the full prompt writing guide, see the director skill's prompt-craft.md.

Video Modes

Finished Cut is the default. Use Clip Stock when the user needs individual shots to combine in an editor.

Multi-Segment Basics

When a video exceeds 15s, split into segments:

1. Each segment 5–15s (default 15s unless beat alignment or pacing requires otherwise)
2. Repeat full character description in every segment prompt
3. Start each segment after S1 with "Continuing from the previous shot: [ending state of prev segment]"
4. Choose the continuity method for the next segment:

- Exact carried-over opening state → extract the previous segment tail frame with ffmpeg, upload it, use ID:first_frame - Motion/style carryover → use task chain <id> to turn a completed segment into ref_video material for the next

Tail-frame → next first-frame

ffmpeg -sseof -0.2 -i generated/shots/S1.mp4 -frames:v 1 -q:v 2 -y generated/keyframes/S1-end.jpg
node ${CLAUDE_SKILL_DIR}/renoise-cli.mjs material upload generated/keyframes/S1-end.jpg
# Use returned material ID as S2 first_frame
node ${CLAUDE_SKILL_DIR}/renoise-cli.mjs task generate \
  --prompt "Continuing from the previous shot: <S2 prompt>" \
  --duration 15 --ratio 16:9 --materials "ID:first_frame"

ref_video chaining

node ${CLAUDE_SKILL_DIR}/renoise-cli.mjs task chain <id>
# Use returned material ID as the next segment's ref_video
node ${CLAUDE_SKILL_DIR}/renoise-cli.mjs task generate \
  --prompt "Continuing from the previous shot: <S2 prompt>" \
  --duration 15 --ratio 16:9 --materials "ID:ref_video"

Face handling decision tree

Never pass raw face images as ref_image — privacy detection will block them.

Quick Templates

Pick the model by job: nano-banana-2 for drafts / banners / wide aspect ratios, nano-banana-pro for hero / final frames, gpt-image-2 for anything with text/logos/typography, midjourney-v7 for stylized illustration (no --resolution).

Product Design Sheet

node ${CLAUDE_SKILL_DIR}/renoise-cli.mjs task generate \
  --prompt "Professional product design sheet showing a sleek wireless headphone from 6 angles: front view, side view, back view, top view, 3/4 perspective, detail close-up. Clean white background, studio lighting, consistent shadow direction." \
  --model nano-banana-2 --resolution 2k --ratio 1:1

Scene / Background Image

node ${CLAUDE_SKILL_DIR}/renoise-cli.mjs task generate \
  --prompt "A modern minimalist living room at golden hour, floor-to-ceiling windows overlooking a city skyline, warm sunlight, photorealistic, 8K detail." \
  --model nano-banana-2 --resolution 2k --ratio 16:9

Hero / Key Frame (higher fidelity)

node ${CLAUDE_SKILL_DIR}/renoise-cli.mjs task generate \
  --prompt "Cinematic hero shot: a lone astronaut standing on a red Martian plain, Earth visible in the dusty sky, volumetric sunlight, photorealistic, highly detailed, shallow depth of field." \
  --model nano-banana-pro --resolution 2k --ratio 16:9

Poster / Graphic with Text

node ${CLAUDE_SKILL_DIR}/renoise-cli.mjs task generate \
  --prompt "Minimalist product poster with bold serif headline 'MIDNIGHT BLOOM' centered, subtitle 'Eau de Parfum' in small caps, dark navy background, gold foil accents." \
  --model gpt-image-2 --resolution 2k --ratio 16:9

Stylized Illustration

node ${CLAUDE_SKILL_DIR}/renoise-cli.mjs task generate \
  --prompt "Stylized fantasy portrait of an elven archer, moody forest lighting, painterly, intricate details." \
  --model midjourney-v7 --ratio 3:4

Error Handling

References

• Video Model Capabilities — Model specs, camera movement reliability, style keywords
• API Endpoint Reference — Raw API endpoints and request/response formats