visual-asset-generator

name	visual-asset-generator
description	Executes AI visual asset generation using prompts from visual-asset-prompt-generator. Calls fal.ai (Nano Banana/Flux 1.1 Pro) for images and Hunyuan3D (via Blender MCP or HuggingFace) for 3D models. Handles async polling and asset post-processing.

Visual Asset Generator

Takes structured prompts from visual-asset-prompt-generator and generates actual visual assets via AI APIs. Manages async 3D generation, image downloads, and asset organization.

When to Use

• After visual-asset-prompt-generator has produced prompts JSON
• User says "generate the images", "create the 3D models", "run the asset pipeline"
• As Step 2 of the brand-visual-pipeline

Input Variables

• [BRAND_NAME] — Brand name (for directory paths)
• [PROMPTS_FILE] — Path to prompts JSON from Skill 1 (default: brands/{BRAND_NAME}/visual-prompts.json)
• [OUTPUT_DIR] — Asset output directory (default: assets/{BRAND_NAME}/)

Required Environment

FAL_KEY=xxx          # fal.ai API key (for Nano Banana / Flux)
HF_TOKEN=hf_xxx     # HuggingFace Pro token (for Hunyuan3D fallback)

The Protocol

Phase 1: Load Prompts

Read the prompts JSON generated by visual-asset-prompt-generator. Group by service:

• fal-flux prompts → fal.ai pipeline
• hunyuan3d prompts → Blender MCP or HuggingFace pipeline
• polyhaven prompts → Blender MCP PolyHaven search

Phase 2: Image Generation via fal.ai

Use scripts/fal_client.py for each image prompt:

python scripts/fal_client.py \
  --prompt "the generation prompt" \
  --negative-prompt "things to avoid" \
  --width 1920 --height 1080 \
  --guidance-scale 7.5 \
  --steps 50 \
  --output "assets/{brand}/images/hero-shots/{asset_id}.png"

fal.ai API Pattern:

import fal_client

# For Flux 1.1 Pro (high quality)
result = fal_client.subscribe(
    "fal-ai/flux-pro/v1.1",
    arguments={
        "prompt": prompt_text,
        "image_size": {"width": width, "height": height},
        "num_inference_steps": steps,
        "guidance_scale": guidance_scale,
        "safety_tolerance": "2"
    }
)
image_url = result["images"][0]["url"]

# For Nano Banana (image cleanup/enhancement)
result = fal_client.subscribe(
    "fal-ai/nano-banana",
    arguments={
        "prompt": prompt_text,
        "image_url": input_image_url,  # if enhancing existing image
        "image_size": {"width": width, "height": height}
    }
)

Post-processing for each generated image:

1. Download from result URL
2. Save as PNG to assets/{brand}/images/{type}/{asset_id}.png
3. Convert to WebP for web use
4. Generate 800px and 400px thumbnails

Phase 3: 3D Model Generation via Hunyuan3D

Option A: fal.ai (preferred — no Blender required)

Use scripts/fal_3d_client.py for each 3D prompt:

# Text-to-3D
python scripts/fal_3d_client.py \
  --mode text-to-3d \
  --prompt "the 3D generation prompt" \
  --enable-pbr \
  --face-count 100000 \
  --output "assets/{brand}/3d-models/{asset_id}/"

# Image-to-3D (from generated product image)
python scripts/fal_3d_client.py \
  --mode image-to-3d \
  --image-url "https://fal.media/files/..." \
  --output "assets/{brand}/3d-models/{asset_id}/"

Output: GLB + OBJ + preview.png in the output directory.

Option B: Blender MCP (when Blender is running)

# Step 1: Check availability
status = mcp__blender__get_hunyuan3d_status(user_prompt="Generate 3D for {brand}")

# Step 2: Generate model
result = mcp__blender__generate_hunyuan3d_model(
    text_prompt=prompt_data["prompt"],
    user_prompt="Generate 3D model for {brand}"
)
# Returns: {"job_id": "job_xxx"}

# Step 3: Poll for completion (30-second intervals)
while True:
    status = mcp__blender__poll_hunyuan_job_status(job_id=job_id)
    if status["status"] == "DONE":
        zip_url = status["ResultFile3Ds"]
        break
    elif status["status"] in ["FAILED", "CANCELED"]:
        # Fall back to Option B
        break
    # Wait 30 seconds between polls

# Step 4: Import into Blender
mcp__blender__import_generated_asset_hunyuan(
    name=f"{brand}_{asset_id}",
    zip_file_url=zip_url
)

# Step 5: Capture viewport render
mcp__blender__get_viewport_screenshot(max_size=1920, user_prompt="Render {brand} model")

# Step 6: Export GLB
mcp__blender__execute_blender_code(
    code=f"""
import bpy
bpy.ops.export_scene.gltf(
    filepath='{output_dir}/3d-models/{asset_id}/model.glb',
    export_format='GLB'
)
""",
    user_prompt="Export model"
)

Option C: HuggingFace API (fallback when fal.ai and Blender are unavailable)

Use scripts/hunyuan_generate.py:

python scripts/hunyuan_generate.py \
  --prompt "the 3D generation prompt" \
  --output "assets/{brand}/3d-models/{asset_id}/"

Phase 4: Texture Search via PolyHaven

For texture prompts:

results = mcp__blender__search_polyhaven_assets(
    asset_type="textures",
    categories=prompt_data["categories"],
    user_prompt="Find textures for {brand}"
)

mcp__blender__download_polyhaven_asset(
    asset_id=chosen_id,
    asset_type="textures",
    resolution="4k",
    user_prompt="Download texture"
)

Phase 5: Generate Asset Manifest

Create assets/{BRAND_NAME}/manifest.json:

{
  "brand": "[BRAND_NAME]",
  "generated_at": "ISO_DATE",
  "assets": [
    {
      "asset_id": "hero-001",
      "type": "hero-image",
      "service": "fal-flux",
      "status": "completed",
      "files": {
        "primary": "images/hero-shots/hero-001.png",
        "webp": "images/hero-shots/hero-001.webp",
        "thumb_800": "images/hero-shots/hero-001-800.png",
        "thumb_400": "images/hero-shots/hero-001-400.png"
      },
      "dimensions": { "width": 1920,   
       
        ... 
    
    
       
       
       
       
       
         
         
      
       
    
  

Error Handling

Error	Action
fal.ai 401	Check FAL_KEY env var, report to user
fal.ai 429	Retry with exponential backoff (max 3 retries)
Hunyuan3D timeout (>10 min)	Fall back to HF API or report
Blender not connected	Use HF API fallback
Image download fails	Retry once, then skip with warning

Scripts

scripts/fal_client.py

fal.ai API client for Flux 1.1 Pro and Nano Banana image generation.

scripts/fal_3d_client.py

fal.ai 3D model generation client. Supports four modes:

# Text to 3D (Hunyuan3D V3) — $0.375/gen, ~2 min, PBR materials
python scripts/fal_3d_client.py --mode text-to-3d --prompt "a leather journal" --output ./output/

# Image to 3D (Hunyuan3D V3) — $0.375/gen, ~2 min
python scripts/fal_3d_client.py --mode image-to-3d --image-url "https://..." --output ./output/

# Image to 3D (TripoSR) — $0.07/gen, <1s, lower quality
python scripts/fal_3d_client.py --mode triposr --image-url "https://..." --output ./output/

# Image to 3D (Hunyuan3D V2) — $0.16/gen
python scripts/fal_3d_client.py --mode hunyuan-v2 --image-url "https://..." --output ./output/

Hunyuan3D V3 extra options:

• --enable-pbr — PBR materials (default: true)
• --face-count 100000 — Target polygon count
• --generate-type Normal|LowPoly|Geometry

Output: GLB, OBJ, and preview PNG files in the specified output directory.

scripts/hunyuan_generate.py

HuggingFace API client for Hunyuan3D when Blender MCP and fal.ai are unavailable.

Integration

• Upstream: visual-asset-prompt-generator (consumes prompts JSON)
• Downstream: visual-asset-integrator (consumes asset manifest)
• External APIs: fal.ai, HuggingFace, Blender MCP
• Part of: brand-visual-pipeline orchestrator