Run any Skill in Manus with one click

scope-image-orchestrator

Stars7

Forks0

UpdatedJune 14, 2026 at 04:48

Use when an image-generation workflow needs SCOPE-style semantic decomposition, prompt optimization, API routing, visual verification, repair loops, batch regression, or when the user starts command mode with 生图优化.

Installation

Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.

Run Skill in Manus

Source

TokinoChanshi

TokinoChanshi/scope-image-orchestrator

View GitHub Repository View Creator Repositories

Download

Run Skill in Manus

Related occupationsSOC

Based on SOC occupation classification

Software DevelopersComputer and Mathematical Occupations·SOC 15-1252

SKILL.md

readonly

Run any Skill with one click

name	scope-image-orchestrator
description	Use when an image-generation workflow needs SCOPE-style semantic decomposition, prompt optimization, API routing, visual verification, repair loops, batch regression, or when the user starts command mode with 生图优化.

SCOPE Image Orchestrator

Goal

SCOPE Image Orchestrator is a Codex skill for image-generation orchestration. It applies the SCOPE paper's core idea: preserve important user requirements as explicit semantic commitments, then keep generation, verification, and repair aligned to those commitments.

This skill is model-agnostic and endpoint-agnostic, but API wire formats must be explicit. Public examples should use official OpenAI and Google Gemini request shapes and should name models only, not deployment-channel labels or relay-specific wording. Configure endpoints through environment variables or a task-local config file.

Command Mode

生图优化 is the human-facing activation command. When the user says it, enter command mode for the current thread. After that, short follow-up messages should be interpreted as image commands until the user says 退出生图优化, 停止生图优化, or clearly switches topics.

Command mode is natural-language first: infer the prompt, count, route, reference image, env files, and output directory from the current thread when safe. Ask only for missing critical items, such as an env file or reference image path.

Command	Meaning	Default implementation
`帮助` / `命令`	Show available command-mode aliases.	`python scripts/scope_commands.py commands`
`查看预设`	List all route presets from the unified library.	`python scripts/scope_commands.py list-presets --detail`
`查看预设 bathroom`	List one route preset.	`python scripts/scope_commands.py list-presets --route bathroom --detail`
`查看主题包`	List all image theme packs from the unified library.	`python scripts/scope_commands.py list-theme-packs --detail`
`查看主题包 mecha_tokusatsu`	List one image theme pack.	`python scripts/scope_commands.py list-theme-packs --theme-pack mecha_tokusatsu --detail`
`跑图 <prompt>` / `单张跑 <prompt>`	Generate one image with the v2 router.	`python scripts/generate_single_v2.py --env-file <env> --user-prompt "<prompt>" --out-dir <out>`
`批量跑 N 张 <prompt>`	Run the same prompt N times; count comes from the user.	`python scripts/scope_commands.py batch-run --count N --env-file <env> --user-prompt "<prompt>" --out-dir <out>`
`视频创作 <prompt>` / `跑视频 <prompt>`	Generate one clip from a route-aware video plan.	`python scripts/scope_commands.py video-run --env-file <env> --user-prompt "<prompt>" --out-dir <out>`
`批量跑视频 N <prompt>`	Run the same video request N times; count from user.	`python scripts/scope_commands.py video-batch --count N --env-file <env> --user-prompt "<prompt>" --out-dir <out>`
`视频分镜创作 <prompt>` / `创作 <prompt> 视频`	Storyboard extraction + per-shot candidates + scoring + selection + assembly (when URLs exist). If 目标时长和`--shot-duration`给出，将按时长切片成多个镜头，默认每镜可自动提候选并打分。	`python scripts/scope_commands.py video-story --env-file <env> --user-prompt "<prompt>" --out-dir <out>`
`视频分镜创作 <prompt> [--send]`	默认是 dry-run；补充 `--send` 才会真实调视频 API。	`python scripts/scope_commands.py video-story --env-file <env> --user-prompt "<prompt>" --out-dir <out> --send`
`视频分镜创作 <prompt> --interactive`	人工逐镜挑选（按候选序号选择）	`python scripts/scope_commands.py video-story --env-file <env> --user-prompt "<prompt>" --out-dir <out> --interactive`
`参考生图 <image> <prompt>`	Generate using a reference image for style/composition/identity/product guidance.	`python scripts/scope_commands.py reference-run --reference-image <image> --env-file <env> --user-prompt "<prompt>" --out-dir <out>`
`严格链路 <prompt>`	Use the paper-style decomposition/synthesis/coverage chain.	Use the strict SCOPE runner for decomposition, synthesis, coverage verification, generation, and repair.
`一句话视频创作 <prompt>`	Natural-language video creation. Parses total duration/shot duration/candidates and runs the storyboard candidate-selection flow automatically.	`python scripts/run_video_skill.py --env-file <env> --user-input "<prompt>" --out-dir <out>`
`回归测试` / `预设回归`	Run route regression after preset or model changes.	`python scripts/run_v2_route_regression.py --env-file <env> --out-dir <out>`
`审核 <image_root>`	Run visual audit over generated images.	`python scripts/audit_generated_images_with_vision.py --env-file <vision.env> --image-root <image_root> --out-file <out.json>`
`三模型对比 <image_root> [--vision-models]`	Run one root against Gemini/Claude/GPT model list (comma-separated).	`python scripts/audit_generated_images_with_vision.py --env-file <vision.env> --image-root <image_root> --vision-models "gemini-3.5-flash,claude-3.5-sonnet,gpt-5.5" --out-file <out.json>`

Reference-image behavior:

By default, 参考生图 uses a vision model to analyze the reference image, writes reference_image.json, injects a compact reference brief into the prompt, then uses the normal image generator. This works even when the image API is text-to-image only.
For direct official reference-image calls, use SCOPE_IMAGE_FORMAT=openai-responses-image or SCOPE_IMAGE_FORMAT=google-gemini-image; the runner will attach the image using input_image or Gemini inlineData.
The OpenAI Images API reference workflow uses /v1/images/edits multipart. Treat it as a separate implementation path, not the default JSON generation path.
Use --reference-mode style|composition|identity|character|product|auto to control what the reference should influence.

Paper Basis

This is a practical implementation inspired by the SCOPE paper, not an official reproduction of the authors' code.

Paper reference:

Title: SCOPE: Structured Decomposition and Conditional Skill Orchestration for Complex Image Generation.
arXiv: https://arxiv.org/abs/2605.08043
Project page: https://nopnor.github.io/SCOPE/

Key ideas adapted here:

Semantic commitments: represent user requirements as persistent objects, attributes, relations, layout rules, text requirements, and factual constraints.
Object-centric decomposition: split a request into entities, constraints, and unknowns before prompting an image model.
Conditional correction: resolve only the missing or ambiguous pieces that matter for the current generation.
Prompt synthesis from commitments: synthesize a compact production prompt from the structured specification instead of relying on a flat prompt alone.
Itemized verification: verify generated outputs entity-by-entity and constraint-by-constraint; do not rely only on a holistic score.
Targeted repair: map failures to retrieval, prompt rewrite, regeneration, or image editing.
Regression evaluation: test routes and presets across multiple prompt families to detect prompt-library or model-routing regressions.

Model names mentioned by the original paper and this implementation may include text-model, image-model, vision-model, text-model, image-model, and vision-model. Public docs should list model names only and avoid private endpoint or relay names.

Read references/scope-paper.md for a longer paper-to-implementation mapping.

Preset Library Notice

references/scope-preset-library.json is the single preset entrypoint. It contains runtime route presets, optimizer guides, and distilled prompt-pattern notes.

Most preset ideas were derived from public internet prompt libraries and prompt examples, then restructured, filtered, rewritten, and distilled into route-level controls. The preset library is not intended to be a verbatim copy of third-party prompts. Keep only compact patterns, route hints, material/camera/layout controls, negative constraints, and source/count metadata.

Current routes:

portrait, magazine, poster, cosplay, interior, product, bathroom,
idiom_cinema, documentary, strategy_overhead, anime_cel

Quick Workflow

Choose an output directory, e.g. scope_runs/<short-task-name>/.
Decompose the request into:
- entities: people, characters, objects, text blocks, places, styles, logos, UI elements.
- constraints: atomic attribute, relation, layout, style, text, or factual requirements.
- unknowns: missing facts or reference details tied to an entity, constraint, or prompt.
Resolve unknowns only when needed using retrieval, reasoning, or a vision/text model.
Route the request to a preset family: portrait, magazine, poster, cosplay, interior, product, bathroom, idiom_cinema, documentary, strategy_overhead, or anime_cel.
Optimize the prompt with a text model such as text-model or vision-model.
Generate the image with an image model such as image-model.
Verify item by item with a vision-capable model such as vision-model.
Repair targeted failures through prompt rewrite, regeneration, or image editing when available.
Stop when critical requirements pass or when additional attempts are unlikely to help.

Model Roles

Role	Example model names	Notes
`decomposer`	`text-model`, `vision-model`	Convert user request into structured commitments.
`prompt_optimizer`	`text-model`, `vision-model`	Turn route + commitments into a compact generation prompt.
`retrieval_corrector`	`vision-model`	Clean factual/reference details when retrieval is available.
`image_generator`	`image-model`, `image-model`	Produce the image.
`verifier`	`vision-model`, `vision-model`	Inspect generated outputs item by item.
`repair_planner`	`text-model`, `vision-model`	Convert failures into prompt repairs or rerun strategy.

Environment Contract

Prefer generic SCOPE_* variables in examples. Set the adapter format first:

SCOPE_LLM_FORMAT=openai-responses
SCOPE_VISION_FORMAT=openai-responses
SCOPE_IMAGE_FORMAT=openai-images

Supported formats:

openai-chat             -> OpenAI Chat Completions
openai-responses        -> OpenAI Responses
google-gemini           -> Gemini generateContent
openai-images           -> OpenAI Images API
openai-responses-image  -> OpenAI Responses image_generation tool
google-gemini-image     -> Gemini native image generation
generic-text-json       -> generic JSON text wrapper
generic-vision-json     -> generic JSON vision wrapper
generic-image-json      -> generic JSON image wrapper
openai-images-legacy    -> legacy OpenAI-compatible JSON image endpoints

Default OpenAI-style example:

SCOPE_LLM_BASE_URL=https://api.openai.com/v1
SCOPE_LLM_API_KEY=your_key
SCOPE_LLM_MODEL=text-model

SCOPE_VISION_BASE_URL=https://api.openai.com/v1
SCOPE_VISION_API_KEY=your_key
SCOPE_VISION_MODEL=text-model

SCOPE_IMAGE_BASE_URL=https://api.openai.com/v1
SCOPE_IMAGE_API_KEY=your_key
SCOPE_IMAGE_MODEL=image-model

Gemini-style example:

SCOPE_LLM_FORMAT=google-gemini
SCOPE_VISION_FORMAT=google-gemini
SCOPE_IMAGE_FORMAT=google-gemini-image
SCOPE_GOOGLE_API_KEY_AUTH=header
SCOPE_LLM_BASE_URL=https://generativelanguage.googleapis.com/v1beta
SCOPE_VISION_BASE_URL=https://generativelanguage.googleapis.com/v1beta
SCOPE_IMAGE_BASE_URL=https://generativelanguage.googleapis.com/v1beta

Read references/api-providers.md for exact request/response shapes.

Use generic adapters when the user has a custom wrapper endpoint:

SCOPE_LLM_FORMAT=generic-text-json
SCOPE_LLM_ENDPOINT_URL=https://example.com/text
SCOPE_IMAGE_FORMAT=generic-image-json
SCOPE_IMAGE_ENDPOINT_URL=https://example.com/image
SCOPE_GENERIC_AUTH_MODE=bearer

Use openai-images-legacy when the user needs the previous image API format:

SCOPE_LLM_FORMAT=openai-chat
SCOPE_VISION_FORMAT=openai-chat
SCOPE_IMAGE_FORMAT=openai-images-legacy
SCOPE_IMAGE_GENERATIONS_URL=https://example.com/v1/images/generations
SCOPE_RESPONSE_FORMATS=b64_json,url,b64_json,url

Legacy direct reference fields are allowed only when explicitly configured:

SCOPE_SEND_REFERENCE_IMAGE=1
SCOPE_REFERENCE_IMAGE_FIELD=images

Output Artifacts

A serious run should persist compact artifacts:

scope_runs/<task>/
  user_request.txt
  route.json
  video_route.json
  optimized_prompt.json
  story_plan.json
  shot_candidates.json
  shot_selection.json
  assembly_plan.json
  assembly_result.json
  assembly/final_video.mp4 (when assembly succeeds)
  video_story_final_summary.json
  video_final_summary.json
  generation_prompt.txt
  generation_prompt.attempt_N.txt
  image_result.attempt_N.json
  visual_audit.attempt_N.json
  video_response.attempt_N.json
  reference_image.json
  image.png
  final_summary.json

Practical Commands

Dry-run a single request:

python scripts/generate_single_v2.py --env-file <env> --user-prompt "生活方式服饰室内场景室内生活方式" --out-dir scope_runs/single_bathroom --dry-run

Run one request with optimizer and visual audit:

python scripts/generate_single_v2.py --env-file <image.env> --llm-env-file <llm.env> --vision-env-file <vision.env> --llm-model text-model --vision-model vision-model --image-model image-model --user-prompt "cinematic poster concept" --out-dir scope_runs/single_poster --max-generation-attempts 2

Run a prompt multiple times:

python scripts/scope_commands.py batch-run --count 6 --env-file <image.env> --user-prompt "commercial product image" --out-dir scope_runs/batch_product

List presets:

python scripts/scope_commands.py list-presets --detail

Validation

When changing presets, model routing, or retry logic:

python scripts/run_v2_route_regression.py --env-file <image.env> --llm-env-file <llm.env> --vision-env-file <vision.env> --out-dir scope_runs/scope_v2_regression

For a small validation run:

python scripts/run_v2_route_regression.py --env-file <image.env> --max-cases 3 --skip-vision --dry-run

name	scope-image-orchestrator
description	Use when an image-generation workflow needs SCOPE-style semantic decomposition, prompt optimization, API routing, visual verification, repair loops, batch regression, or when the user starts command mode with 生图优化.