Media Gen

Media Gen 🎬

Generate images and videos with a single AIsa API key. Full support for every image and video model AIsa routes through its Unified LLM Gateway, across three different endpoint paths.

Compatibility

Works with any agentskills.io-compatible harness, including:

• Claude Code and Claude (Anthropic)
• OpenAI Codex
• Cursor
• Gemini CLI (Google)
• OpenCode, Goose, OpenClaw, Hermes
• and any other harness that implements the [Agent Skills

specification](https://agentskills.io/specification)

Requires Python 3, a POSIX shell, and AISA_API_KEY (get one at aisa.one).

🔥 What You Can Do

Image — Gemini (base64 inline)

"Generate a cyberpunk-style city nightscape, neon lights, rainy night, cinematic feel"

Image — Wan 2.7 (URL in chat response)

"Generate an ultra-detailed product shot of a red panda, studio lighting, sharp focus"

Image — Seedream (OpenAI-compatible, large format)

"Generate a 2048×2048 magazine cover: neo-noir detective portrait, film grain"

Video — text-to-video (Wan t2v)

"Sweeping establishing shot of a neon cyberpunk skyline at dusk, 5 seconds"

Video — image-to-video (Wan i2v)

"Starting from this reference image, gentle camera push-in with parallax"

Supported Models

Image generation — 4 models, 3 endpoints

| Model | Developer | Endpoint | Notes | |---|---|---|---| | gemini-3-pro-image-preview | Google | POST /v1/models/{model}:generateContent | Images returned as base64 in candidates[].parts[].inline_data | | wan2.7-image | Alibaba | POST /v1/chat/completions | Images returned as URL parts in choices[].message.content[] (type=image). $0.030/image | | wan2.7-image-pro | Alibaba | POST /v1/chat/completions | Higher fidelity. $0.075/image | | seedream-4-5-251128 | ByteDance | POST /v1/images/generations | OpenAI-compatible. Minimum 3,686,400 pixels (e.g. 1920×1920). $0.040/image |

Video generation — 4 Wan variants, 1 endpoint

| Model | Kind | Image field | Output SR | |---|---|---|---| | wan2.6-t2v | text-to-video | none | 1080 | | wan2.6-i2v | image-to-video | input.img_url (string) | 720 | | wan2.7-t2v | text-to-video | none | 720 | | wan2.7-i2v | image-to-video | input.media (array) ⚠ | 720 |

> ⚠ Schema trap on wan2.7-i2v. It takes the reference image in > input.media (array of URLs), not input.img_url like > wan2.6-i2v. Submissions without media return HTTP 200 with a > task_id, then fail downstream with InvalidParameter: Field required: > input.media. The bundled client routes this automatically — just > pass --img-url and pick the model.

Quick Start

export AISA_API_KEY="your-key"

# Any image model — client routes to the right endpoint
python3 {baseDir}/scripts/media_gen_client.py image \
  --model gemini-3-pro-image-preview \
  --prompt "A cute red panda, cinematic lighting" \
  --out out.png

python3 {baseDir}/scripts/media_gen_client.py image \
  --model wan2.7-image-pro \
  --prompt "Ultra-detailed product shot of a red panda" \
  --out out.png

python3 {baseDir}/scripts/media_gen_client.py image \
  --model seedream-4-5-251128 \
  --prompt "Neo-noir detective portrait, film grain" \
  --size 2048x2048 \
  --out out.png

# Video — text-to-video (no image needed)
python3 {baseDir}/scripts/media_gen_client.py video-create \
  --model wan2.7-t2v \
  --prompt "Sweeping shot of a neon cyberpunk skyline"

# Video — image-to-video on wan2.7-i2v (client routes to input.media[])
python3 {baseDir}/scripts/media_gen_client.py video-create \
  --model wan2.7-i2v \
  --prompt "gentle zoom with parallax" \
  --img-url "https://example.com/reference.jpg" \
  --duration 5

# Wait and download
python3 {baseDir}/scripts/media_gen_client.py video-wait \
  --task-id  --download --out out.mp4

---

🖼️ Image Generation — endpoint reference

Gemini family → POST /v1/models/{model}:generateContent

Documentation: Google Gemini Chat.

curl -X POST "https://api.aisa.one/v1/models/gemini-3-pro-image-preview:generateContent" \
  -H "Authorization: Bearer $AISA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "contents":[
      {"role":"user","parts":[{"text":"A cute red panda, cinematic lighting"}]}
    ]
  }'

Response contains candidates[].parts[].inline_data with {mime_type, data} where data is a base64 PNG.

Wan 2.7 family → POST /v1/chat/completions

Documentation: Image Generation via Chat.

Critical rule: messages[].content must be an array of typed parts. A plain string returns HTTP 400 invalid_parameter_error.

curl -X POST "https://api.aisa.one/v1/chat/completions" \
  -H "Authorization: Bearer $AISA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "wan2.7-image",
    "messages": [
      {"role":"user","content":[
        {"type":"text","text":"A cute red panda, ultra-detailed, cinematic lighting"}
      ]}
    ],
    "n": 1
  }'

Images come back as {type: "image", image: ""} parts inside choices[].message.content[].

Seedream → POST /v1/images/generations

Documentation: OpenAI-Compatible Image Generations.

curl -X POST "https://api.aisa.one/v1/images/generations" \
  -H "Authorization: Bearer $AISA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "seedream-4-5-251128",
    "prompt": "A cute red panda, ultra-detailed, cinematic lighting",
    "n": 1,
    "size": "2048x2048"
  }'

Response: data[].url or data[].b64_json. Upstream enforces a minimum of 3,686,400 pixels. 1024×1024 and 1536×1536 get rejected. Any aspect ratio works as long as width × height ≥ 3,686,400.

---

🎞️ Video Generation — endpoint reference

Create task → POST /apis/v1/services/aigc/video-generation/video-synthesis

Documentation: Create video generation task. Header X-DashScope-Async: enable is required.

# wan2.6-t2v — text-to-video
curl -X POST "https://api.aisa.one/apis/v1/services/aigc/video-generation/video-synthesis" \
  -H "Authorization: Bearer $AISA_API_KEY" \
  -H "Content-Type: application/json" \
  -H "X-DashScope-Async: enable" \
  -d '{
    "model":"wan2.6-t2v",
    "input":{"prompt":"cinematic close-up, slow push-in"},
    "parameters":{"resolution":"720P","duration":5}
  }'

# wan2.7-i2v — image-to-video (⚠ input.media not input.img_url)
curl -X POST "https://api.aisa.one/apis/v1/services/aigc/video-generation/video-synthesis" \
  -H "Authorization: Bearer $AISA_API_KEY" \
  -H "Content-Type: application/json" \
  -H "X-DashScope-Async: enable" \
  -d '{
    "model":"wan2.7-i2v",
    "input":{
      "prompt":"gentle zoom with parallax",
      "media":["https://example.com/reference.jpg"]
    },
    "parameters":{"resolution":"720P","duration":5}
  }'

Poll task → GET /apis/v1/services/aigc/tasks/{task_id}

Documentation: Get video generation task result.

> task_id is a path parameter. The query-string form > ?task_id=... returns HTTP 500 unsupported uri.

curl "https://api.aisa.one/apis/v1/services/aigc/tasks/YOUR_TASK_ID" \
  -H "Authorization: Bearer $AISA_API_KEY"

---

Python Client

The bundled client at scripts/media_gen_client.py auto-routes each image model to the correct endpoint and normalizes the response to a saved file.

# Image — model picks the endpoint
python3 {baseDir}/scripts/media_gen_client.py image \
  --model  \
  --prompt "..." \
  --out out.png

# Video — create task
python3 {baseDir}/scripts/media_gen_client.py video-create \
  --model  \
  --prompt "..." \
  [--img-url https://... (required for -i2v models)] \
  [--duration 5|10] \
  [--resolution 720P|1080P]

# Video — poll / wait / download
python3 {baseDir}/scripts/media_gen_client.py video-status --task-id 
python3 {baseDir}/scripts/media_gen_client.py video-wait --task-id  --poll 10 --timeout 600
python3 {baseDir}/scripts/media_gen_client.py video-wait --task-id  --download --out out.mp4

API Reference

This skill calls the following AIsa endpoints directly:

• Google Gemini Chat — generateContent — Gemini image models
• Image Generation via Chat — Wan 2.7 image family
• OpenAI-Compatible Image Generations — Seedream
• Create video generation task — all 4 Wan video variants
• Get video generation task result — async polling

See the full AIsa API Reference for the complete catalog.

License

MIT — see [LICENSE](../LICENSE) at the repo root.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: AIsa-team
• Source: AIsa-team/agent-skills
• License: Apache-2.0

Install and usage instructions live in the source repository linked above.