AI Agent: batch asset factory — source folder → validated 3D asset pack

[fernandotonon]

Child of #1000. Integration follow-up to #1007.

Goal

Allow the AI agent to process a folder of source images and autonomously create a batch of validated 3D assets for a game or project.

Example request:

“Process /SchoolGame/assets/source. Create one 3D asset per image. Characters should be game-ready, rigged, skinned, low-poly, and receive idle/walk/run animations. Props should be optimized, textured, and exported as GLB.”

Scope

Support a project-level workflow where the agent:

1. scans a folder of input images;
2. classifies each source as character, prop, environment object, creature/monster, or uncertain;
3. creates a per-asset task plan;
4. runs the appropriate workflow for each asset;
5. stores intermediate artifacts and final outputs;
6. validates each asset;
7. retries or escalates uncertain cases;
8. reports a final batch summary.

Input discovery

• Accept a source folder path.
• Enumerate supported image files.
• Create a persistent asset manifest/job file for resumable execution.
• Detect duplicate or very similar inputs when feasible.

Asset classification

For each input image, classify into categories such as:

• humanoid character
• creature/monster character
• rigid prop
• environment/static object
• uncertain/manual review

The agent should choose the pipeline accordingly.

Character pipeline

Typical steps:

• inspect input-image suitability;
• remove background / clean up if needed;
• optionally use the existing FLUX.2 image-generation path to regenerate a cleaner full-body neutral reference when the original is unsuitable;
• verify reference suitability when a VLM is available;
• run the configured image-to-3D backend;
• validate geometry;
• simplify/repair if needed;
• auto-rig;
• skin;
• apply selected animation templates (idle/walk/run by default or project preset);
• render a screenshot and optionally run final visual QA;
• export final asset.

Prop/object pipeline

Typical steps:

• inspect input-image suitability;
• optional cleanup/background removal;
• run image-to-3D backend;
• validate geometry;
• simplify/repair if needed;
• generate/apply materials or PBR processing when appropriate;
• render screenshot and optionally run final visual QA;
• export final asset.

Asset manifest / resumability

Add a persistent batch-job format, e.g.:

{
  "project": "school-game",
  "sourceFolder": "/SchoolGame/assets/source",
  "outputFolder": "/SchoolGame/assets/generated",
  "assets": [
    {
      "source": "characters/teacher.jpg",
      "type": "character",
      "status": "completed",
      "requirements": {
        "maxTriangles": 20000,
        "rig": true,
        "animations": ["idle", "walk", "run"]
      },
      "artifacts": {
        "referenceImage": "teacher_reference.png",
        "rawMesh": "teacher_raw.glb",
        "finalMesh": "teacher_final.glb"
      },
      "warnings": []
    }
  ]
}

Requirements:

• resumable after interruption/crash;
• per-asset progress/state;
• cached intermediate artifacts;
• completed steps are not recomputed unless inputs/requirements invalidate them.

Confidence and review gates

Do not blindly accept every result.

The agent should auto-accept high-confidence results and mark uncertain ones as needs_review.

Possible review triggers:

• identity drift on important characters;
• poor facial similarity;
• cropped/ambiguous source;
• failed rigging/skinning validation;
• poor final visual QA;
• repeated unsuccessful retries.

One problematic asset must not stop the entire batch.

Retry policy

• Configurable retry budget per asset and per step.
• Example: up to 2 automatic retries for reference generation or image-to-3D.
• Stop escalating once retries are exhausted and mark the asset for review.

Outputs

• final exported assets in a target folder;
• per-asset screenshots/previews;
• manifest with complete execution status;
• final batch summary, e.g.:

37 source images processed
31 completed automatically
4 completed with warnings
2 require review

UX / surfaces

GUI

Add a “Batch asset factory” workflow/panel:

• choose source folder and output folder;
• presets for character triangle budget, prop triangle budget, required animations, style hints, retry count and trusted/auto-approve mode;
• queue/progress per asset;
• filter completed / warning / review-required assets;
• resume previous batch job.

CLI

Example:

qtmesh batch-assets /SchoolGame/assets/source \
  --output /SchoolGame/assets/generated \
  --character-max-tris 20000 \
  --prop-max-tris 10000 \
  --character-animations idle,walk,run \
  --retry 2

MCP

Add a batch-oriented workflow entry point, e.g. batch_generate_assets / run_asset_pipeline, with progress/job-state queries rather than one enormous blocking tool result.

Dependencies / related work

Depends primarily on:

• AI Agent: introduce AIAgentManager task state machine and structured observations #1001 task state machine
• AI Agent: capability registry and dynamic MCP/tool routing #1002 capability routing
• AI Agent: constrained tool-call schema and typed planner/executor protocol #1003 constrained agent protocol
• AI Agent: local VLM integration for viewport and generated-image inspection #1004 VLM inspection (optional but strongly improves source/reference/final QA)
• AI Agent: model lifecycle and memory-budget manager for orchestrator + specialist models #1005 model lifecycle / memory management
• AI Agent: validation/verifier layer for geometry, rigging, materials and visual QA #1006 verifier layer
• AI Agent: end-to-end FLUX.2 → image-to-3D → rig/skin/animate workflow and agent benchmark #1007 single-asset end-to-end FLUX.2 → 3D → rig/skin/animate workflow

Acceptance

• The agent can process a folder containing mixed characters and props.
• Each asset gets a structured per-asset plan and status.
• Character assets can complete reference prep → image-to-3D → validate → rig → skin → animate → export.
• Prop assets can complete reference prep → image-to-3D → validate → optimize → material/texture processing → export.
• Failed or uncertain assets do not stop the whole batch.
• Interrupted jobs can resume from saved manifest state.
• Cached intermediate artifacts prevent unnecessary recomputation.
• Final batch summary reports completed/warning/review-required assets.
• Text-only mode works in degraded form; VLM mode adds visual source/reference/final checks.

Parent: #1000