image-factory-plugin

Image Factory

Plugin marketplaces

This plugin belongs to AIGC content creation.

Category · Marketplace · Purpose Full-stack development · Full Stack Plugins · Architecture and UI design, code understanding, quality checks, code review, workflow governance, and server operations AIGC content creation · Full AIGC Plugins · Image, video, audio, music, 3D, and multimodal content creation

Turn a reference-driven image task into an auditable production run: a validated plan, an approved spend, and a hash-verified receipt for every artifact.

English | 简体中文 · Install · Quick start · Skills · Troubleshooting

Positioning

image-factory produces verifiable image batches. It turns references and a visual direction into a validated plan, quotes the batch, waits for your approval, generates one image per plan item through the host's image tool, and records a recomputed hash, size, and dimension receipt for every file.

The plugin exposes two layers. In Codex the built-in imagegen / image_gen path is preferred and needs no provider key. The unchanged Baoyu workflows are offered only after an explicitly observed image-quota exhaustion in Codex, or on a host without a verified native image capability. The Factory loop adds the governed batch: plan validation, approval, receipts, evaluation, and recovery.

Property · Value Plugin ID · image-factory Version · 0.9.1+codex.20261018 Maturity · No production-readiness claim; the newest runtime record covers the 0.8.0 candidate Runtime · Python 3.11+, standard library only MCP server · None; the manifest declares no MCP entry

0.8.0 added multi-hash evidence, stratified calibration, a review workspace, and a runtime acceptance matrix. The paid 4/8/12, cross-host, and fault-injection cases remain NOT_RUN; synthetic tests do not establish live character, prop, or style stability.

Installation

Prerequisites

• Python 3.11 or newer on PATH.
• The Codex CLI reachable on PATH, or CODEX_HOME pointing at an installation.
• Factory-only batches need no API key and no npm package.
• The Baoyu fallback requires bun or npx plus whatever credential the selected provider documents; the codex-cli provider reuses an already signed-in Codex CLI.

From the plugin marketplace

codex plugin marketplace add partme-ai/full-aigc-plugins
codex plugin add image-factory@full-aigc-plugins

From the repository-local marketplace

codex plugin marketplace add full-aigc-plugins/image-factory-plugin --ref v0.9.1
codex plugin add image-factory@partme-ai-image-factory

Restart Codex or the ChatGPT desktop app and open a new task so the skills load.

For development, clone the repository and run bin/image-factory probe; it prints a JSON capability report for Python, the Codex CLI, and ledger access and never prints a secret value. codex plugin list should show image-factory@partme-ai-image-factory installed, enabled.

Quick start

1. Find an attributed prompt direction offline:

bin/image-factory prompt-search 'storybook illustration' --limit 3 --json

1. Write a batch plan, then validate and quote it. Validation rejects unknown fields, cap violations, and duplicate idempotency keys before anything is spent:

bin/image-factory validate-plan image-plan.json
bin/image-factory quote image-plan.json

1. Approve the exact plan hash and run:

bin/image-factory run image-plan.json --approval image-approval.json

1. Evaluate or reconcile an interrupted run:

bin/image-factory evaluate image-plan.json
bin/image-factory recover image-plan.json

Each attempt is a fresh streaming subprocess against the built-in image tool. There is no retry loop: a silent retry is how one bad prompt becomes a large bill. A failed item is terminal; optimize produces a new prompt round that needs its own approval.

What you can build

Workflow · Built-in path · Verifiable output Find a direction · prompt-search over bundled templates · Ranked prompt directions with source attribution Produce a governed batch · validate-plan → quote → run · Per-artifact hash, size, and dimension receipts Keep a series consistent · consistency_profile, entity traits, role-aware references · Effective prompt and reference count reported by quote Evaluate and iterate · evaluate + image-factory-review · Deterministic gate results, advisory dimension scores, next-round plan Recover an interrupted run · recover · Reconciled ledger without another generation call Make one ad-hoc image · Codex built-in imagegen · A bitmap asset outside the Factory ledger

Multi-image stories declare consistency_profile; anchors are injected into every selected frame in a fixed order and story_state compiles locks and transitions into the effective prompt. Anchors improve control but cannot guarantee model-level identity.

Skills

Skill · Purpose image-factory-use · Picks the entry point for a batch request and never performs the work itself image-factory-run · Validates, quotes, obtains approval, generates one image per item, and collects receipts image-factory-judge · Runs deterministic gates, records advisory scores beside human labels, and writes the next round image-factory-review · Writes the independent dimensioned advisory document consumed by evaluate and optimize image-factory-recover · Reads the ledger and maps an interrupted or unknown batch to the single legal next step image-factory-harness · Chooses Codex imagegen first, Baoyu only after quota evidence, and normalizes delivery evidence imagegen · Managed snapshot of the Codex raster-image generation and editing instructions baoyu-image-gen · Vendored multi-provider image generation through bun or npx baoyu-cover-image · Vendored article-cover generation across aspect ratios baoyu-xhs-images · Vendored social image-card series generation

The three Baoyu skills come from baoyu-skills v2.5.2, while imagegen and image-factory-use/run/judge/recover come from image-factory-skills v1.1.0; both sets are pinned in skills.lock.json. image-factory-harness and image-factory-review are plugin-local. Edit none of these copies in place.

Slash commands

Command · What it does /image-factory · Total entry point; routes to run, judge, or recover /image-factory-run · Quotes, approves, generates, and collects receipts for a batch /image-factory-judge · Evaluates a finished batch and drives prompt optimization /image-factory-recover · Restores an interrupted batch from the job ledger

Configuration

Setting · Where it lives · Notes Codex home · CODEX_HOME environment variable · Defaults to the standard Codex location Job ledger · The --job path, or .job.json · Plain JSON written atomically Receipts · .receipts/ · One JSON file per artifact Job lock · .lock · Cross-process advisory lock Credentials · None · The plugin stores no secret and the ledger rejects credential-like keys Baoyu provider · Local environment or provider config · Chosen by the user, never written by the plugin

Exit codes are 0 success, 1 failure, 2 usage error, 3 approval required, 4 capability unavailable, 5 job locked, and 6 recovery required.

Architecture and security

flowchart LR
    A[References + direction] --> B[Prompt search]
    B --> C{Validate + quote}
    C -->|invalid| B
    C -->|valid| D{Approval}
    D -->|approved| E[Codex image tool]
    E --> F[Receipts + ledger]
    F --> G{Evaluate}
    G -->|failures| H[Optimize → new approval]
    G -->|none| I[Accepted batch]

Component · Owns · Does not own bin/image-factory · CLI entry point · Business rules scripts/plan_validator.py · Closed plan schema, caps, idempotency keys · Prompt quality scripts/generation_runner.py · One fresh subprocess per attempt · Retry policy scripts/job_ledger.py, scripts/receipt_store.py · Durable state, scrubbing, per-artifact re-verification · Remote lifetime skills/ (10) · Routing, governance, and review instructions · Runtime enforcement

Authentication is Codex's own. The Factory path never calls an external generation API, never updates an upstream dependency on its own, and never writes a credential: the ledger rejects credential-like keys before writing anything. Reference inputs are validated before use, and every hash, size, and dimension is recomputed from disk with a second check that catches a file rewritten mid-validation.

Troubleshooting

Symptom · Check first · Resolution probe reports the Codex CLI missing · PATH or CODEX_HOME · Point CODEX_HOME at the installation or fix PATH Plan validation fails · The plan schema · Fix the named field; unknown fields are rejected on purpose The run refuses to start · The approval file · Quote the plan again and approve the new revision A run stops mid-batch · Ledger state · Run recover; verified receipts are reconciled, not regenerated Hash verification fails · The artifact on disk · Treat it as a hard failure: the file changed after generation The image quota is exhausted · Codex quota evidence · Only then choose a Baoyu fallback and configure its provider locally Generation runs on Linux or Windows · Platform support · The generation path is verified on macOS only; elsewhere it stays NOT_RUN A model name is missing from receipts · Platform boundary · The model is chosen by Codex; receipts record only what Codex reports

Upgrade

codex plugin marketplace upgrade partme-ai-image-factory
codex plugin add image-factory@partme-ai-image-factory

Contributing and support

Open functional issues at . Before proposing a change, state the Python version you verified on, whether it alters the plan schema or the ledger format, and include the affected gates.

License

Apache-2.0 — see LICENSE, NOTICE, and THIRD_PARTY_NOTICES.md. Vendored Baoyu skills and the imagegen snapshot keep their upstream terms.