polym-eval-generate-veo

name	polym-eval-generate-veo
description	Generate videos with Google Veo via Vertex AI or AIDP. Use for Veo text-to-video, image-to-video, and first-frame workflows.

Generate Veo Videos

Generate videos using Google Veo models. Supports two backends:

• vertexai (default): Google Cloud Vertex AI (google-genai SDK)
• aidp: ByteDance AIDP model hub (requests + JSON)

Both T2V (text-to-video) and I2V (image-to-video, first-frame reference) are supported.

---

Pre-flight Check

Vertex AI backend (default)

export VERTEX_PROJECT="your-gcp-project-id"
export VERTEX_LOCATION="us-central1"

AIDP backend

export AIDP_AK="your-aidp-ak"

---

Quick Start

cd skills/polym-eval-generate-veo
pip install -r requirements.txt

Vertex AI backend (default)

# T2V
python scripts/generate.py \
  --prompt "A golden retriever running through a meadow at sunset"

# I2V (first-frame reference)
python scripts/generate.py \
  --prompt "The scene gently comes alive" \
  --first-frame /path/to/photo.jpg

AIDP backend

# T2V
python scripts/generate.py \
  --backend aidp \
  --prompt "A golden retriever running through a meadow at sunset"

# I2V
python scripts/generate.py \
  --backend aidp \
  --prompt "The scene gently comes alive" \
  --first-frame /path/to/photo.jpg

# Use fast model
python scripts/generate.py \
  --backend aidp --model veo-3.0-fast-generate-001 \
  --prompt "A golden retriever running"

---

Parameter Reference

Flag	Default	Description
--prompt / -p	required	Text prompt
--backend	vertexai	vertexai or aidp
--model / -m	veo-3.0-generate-preview (vertexai) / veo-3.0-generate-001 (aidp)	Model override
--first-frame	None	Reference image path or URL (I2V)
--seconds / -d	8	Duration: 4, 6, or 8
--aspect-ratio	auto-detected	16:9 or 9:16
--output / -o	generated/<id>.mp4	Output path

---

Available Models

Backend	Model	Mode	Notes
vertexai	veo-3.0-generate-preview	T2V, I2V	Default Vertex AI model
aidp	veo-3.0-generate-001	T2V, I2V	Default AIDP model, confirmed working
aidp	veo-3.0-fast-generate-001	T2V, I2V	Faster variant, confirmed working

---

Environment Variables

Vertex AI backend

Variable	Required	Description
VERTEX_PROJECT	Yes	GCP project ID
VERTEX_LOCATION	Yes	Vertex AI location (e.g. us-central1)
VEO_MODEL	No	Default model
TOS_BUCKET	No	TOS bucket for network URL
TOS_ACCESS_KEY	No	TOS access key
TOS_SECRET_KEY	No	TOS secret key

AIDP backend

Variable	Required	Description
AIDP_AK	Yes	AIDP API Key (ak)
AIDP_HOST	No	AIDP host (default: aidp.bytedance.net)
VEO_AIDP_MODEL	No	Default AIDP model (default: veo-3.0-generate-001)

---

Output

{
  "gen_success": true,
  "gen_video_path": "generated/abc12345.mp4",
  "gen_url": "https://tosv.byted.org/obj/gpt-openapi-public-cn/xxx_0.mp4",
  "gen_elapsed": 90,
  "gen_failed_reason": null
}

---

AIDP API Details

Submit task

POST /veo/v1/generate?ak={ak}

{
  "model": "veo-3.0-generate-001",
  "instances": [{"prompt": "...", "image": {"bytesBase64Encoded": "...", "mimeType": "image/jpeg"}}],
  "parameters": {"aspectRatio": "16:9", "durationSeconds": 8}
}

Get status

POST /veo/v1/fetch?ak={ak}&product={product_name}

{"model": "veo-3.0-generate-001", "operationName": "projects/.../operations/<id>"}

Response: {"done": true, "response": {"videos": [{"gcsUri": "gpt-openapi-public-cn/xxx_0.mp4"}]}}

Download

https://tosv.byted.org/obj/{gcsUri} (CN internal network)

Model → product name mapping

Model	Product name (for fetch)
veo-3.0-generate-001	google_veo_3.0_generate_001
veo-3.0-fast-generate-001	google_veo_3.0_fast_generate_001

---

Dependencies

requests
Pillow
python-dotenv
google-genai    # vertexai backend only

Optional (for TOS upload / network URL with Vertex AI backend):

tos

---

Troubleshooting

• VERTEX_PROJECT/VERTEX_LOCATION not set: Script will prompt for input (vertexai backend only).
• AIDP_AK not set: Script will prompt for input (aidp backend only).
• TOS download fails: tosv.byted.org requires ByteDance internal network access (VPN or office network).
• google-genai import error: Only needed for vertexai backend — pip install google-genai.
• Unknown product name warning: Model not in mapping; product name auto-derived from model name.