gemini-multimodal

name	gemini-multimodal
description	Generate images (Nano Banana + Imagen 4), videos (Veo), speech (TTS), and music (Lyria) using Gemini API via curl. Zero dependencies. Requires GEMINI_API_KEY (Paid Tier 1+).

Gemini Multimodal Generation

Generate images, videos, speech, and music from Claude Code using the Gemini API. Zero dependencies — uses only curl and python3 (macOS/Linux built-in).

Prerequisites

• GEMINI_API_KEY environment variable set with a Paid Tier 1+ key
• Free tier keys have 0 quota for all generation models
• Get a key at https://aistudio.google.com/apikey — billing must be enabled

When to Use This Skill

• User asks to generate, create, or draw an image
• User asks to generate a video from text or image
• User asks to convert text to speech / generate audio
• User asks to generate music
• User asks to edit an image iteratively

Available Models

Capability	Model ID	Endpoint	Quality
Image gen (fast)	gemini-2.5-flash-image	generateContent	Good, fast
Image gen (pro)	gemini-3-pro-image-preview	generateContent	High quality
Image gen (latest)	gemini-3.1-flash-image-preview	generateContent	Nano Banana 2
Image gen (Imagen)	imagen-4.0-fast-generate-001	predict	Photorealistic
Image gen (Imagen Ultra)	imagen-4.0-ultra-generate-001	predict	Best quality
Video gen	veo-3.1-generate-preview	predictLongRunning	Up to 8s, 4K
TTS	gemini-2.5-flash-preview-tts	generateContent	30 voices
Music	lyria-3-clip-preview	generateContent	30s clips

---

1. Nano Banana Image Generation

Use for general image generation with text-and-image interleaving.

curl -s -X POST "https://generativelanguage.googleapis.com/v1beta/models/gemini-2.5-flash-image:generateContent?key=$GEMINI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "contents":[{"parts":[{"text":"PROMPT_HERE"}]}],
    "generationConfig":{"responseModalities":["TEXT","IMAGE"]}
  }' | python3 -c "
import sys,json,base64
d=json.load(sys.stdin)
if 'error' in d:
    print(f'ERROR: {d[\"error\"][\"message\"]}'); sys.exit(1)
for p in d['candidates'][0]['content']['parts']:
    if 'inlineData' in p:
        img=base64.b64decode(p['inlineData']['data'])
        path='OUTPUT_PATH'
        with open(path,'wb') as f: f.write(img)
        print(f'Saved: {path} ({len(img)} bytes)')
    elif 'text' in p and p['text'].strip():
        print(p['text'])
"

Replace: PROMPT_HERE with the image description, OUTPUT_PATH with the desired file path (e.g., /tmp/output.png).

Model substitution: Replace gemini-2.5-flash-image with gemini-3-pro-image-preview or gemini-3.1-flash-image-preview for different quality/speed tradeoffs.

---

2. Imagen 4 Image Generation

Use for highest quality photorealistic images. Uses the predict endpoint (not generateContent).

curl -s -X POST "https://generativelanguage.googleapis.com/v1beta/models/imagen-4.0-fast-generate-001:predict?key=$GEMINI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "instances":[{"prompt":"PROMPT_HERE"}],
    "parameters":{"sampleCount":1}
  }' | python3 -c "
import sys,json,base64
d=json.load(sys.stdin)
if 'error' in d:
    print(f'ERROR: {d[\"error\"][\"message\"]}'); sys.exit(1)
for i,pred in enumerate(d['predictions']):
    img=base64.b64decode(pred['bytesBase64Encoded'])
    path='OUTPUT_PATH'
    with open(path,'wb') as f: f.write(img)
    print(f'Saved: {path} ({len(img)} bytes)')
"

Parameters:

• sampleCount: 1-4 (number of image variants)
• Model variants: imagen-4.0-fast-generate-001 (fast), imagen-4.0-generate-001 (standard), imagen-4.0-ultra-generate-001 (ultra)

---

3. Veo Video Generation

Asynchronous — submit job, poll for completion, download result.

Step 1: Submit

curl -s -X POST "https://generativelanguage.googleapis.com/v1beta/models/veo-3.1-generate-preview:predictLongRunning?key=$GEMINI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "instances":[{"prompt":"PROMPT_HERE"}],
    "parameters":{"aspectRatio":"16:9","durationSeconds":4}
  }' | python3 -c "
import sys,json
d=json.load(sys.stdin)
if 'error' in d:
    print(f'ERROR: {d[\"error\"][\"message\"]}'); sys.exit(1)
print(d['name'])
"

Save the operation name (e.g., models/veo-3.1-generate-preview/operations/xxxxx).

Parameters:

• durationSeconds: 4, 6, or 8
• aspectRatio: "16:9" or "9:16"

Step 2: Poll

curl -s "https://generativelanguage.googleapis.com/v1beta/OPERATION_NAME?key=$GEMINI_API_KEY" | python3 -c "
import sys,json
d=json.load(sys.stdin)
done=d.get('done',False)
if done and 'response' in d:
    samples=d['response'].get('generateVideoResponse',{}).get('generatedSamples',[])
    for s in samples:
        uri=s.get('video',{}).get('uri','')
        if uri: print(f'READY:{uri}')
elif done and 'error' in d:
    print(f'ERROR: {d[\"error\"]}')
else:
    print('PENDING')
"

Poll every 10-15 seconds until output starts with READY:.

Step 3: Download

curl -s -L "VIDEO_URI&key=$GEMINI_API_KEY" -o OUTPUT_PATH

Replace: VIDEO_URI with the URI from step 2, OUTPUT_PATH with desired path (e.g., /tmp/output.mp4).

---

4. TTS (Text-to-Speech)

Generate speech audio from text. Output is PCM 24kHz mono, saved as WAV.

curl -s -X POST "https://generativelanguage.googleapis.com/v1beta/models/gemini-2.5-flash-preview-tts:generateContent?key=$GEMINI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "contents":[{"parts":[{"text":"Say exactly this: TEXT_HERE"}]}],
    "generationConfig":{
      "responseModalities":["AUDIO"],
      "speechConfig":{"voiceConfig":{"prebuiltVoiceConfig":{"voiceName":"VOICE_NAME"}}}
    }
  }' | python3 -c "
import sys,json,base64
d=json.load(sys.stdin)
if 'error' in d:
    print(f'ERROR: {d[\"error\"][\"message\"]}'); sys.exit(1)
for p in d['candidates'][0]['content']['parts']:
    if 'inlineData' in p:
        audio=base64.b64decode(p['inlineData']['data'])
        path='OUTPUT_PATH'
        with open(path,'wb') as f: f.write(audio)
        print(f'Saved: {path} ({len(audio)} bytes)')
"

Available voices: Zephyr, Puck, Charon, Fenrir, Aoede, Kore, Leda, Orus, Vega, Altair (and 20+ more).

Replace: TEXT_HERE with the text to speak, VOICE_NAME with chosen voice, OUTPUT_PATH with desired path (e.g., /tmp/speech.wav).

Pro model: Replace gemini-2.5-flash-preview-tts with gemini-2.5-pro-preview-tts for higher quality.

---

5. Lyria Music Generation

Generate music clips from text descriptions.

curl -s -X POST "https://generativelanguage.googleapis.com/v1beta/models/lyria-3-clip-preview:generateContent?key=$GEMINI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "contents":[{"parts":[{"text":"PROMPT_HERE"}]}],
    "generationConfig":{"responseModalities":["AUDIO"]}
  }' | python3 -c "
import sys,json,base64
d=json.load(sys.stdin)
if 'error' in d:
    print(f'ERROR: {d[\"error\"][\"message\"]}'); sys.exit(1)
for p in d['candidates'][0]['content']['parts']:
    if 'inlineData' in p:
        audio=base64.b64decode(p['inlineData']['data'])
        path='OUTPUT_PATH'
        with open(path,'wb') as f: f.write(audio)
        print(f'Saved: {path} ({len(audio)} bytes, {p[\"inlineData\"][\"mimeType\"]})')
    elif 'text' in p and p['text'].strip():
        print(p['text'][:200])
"

Replace: PROMPT_HERE with music description (e.g., "A cheerful acoustic guitar melody, upbeat tempo"), OUTPUT_PATH with desired path (e.g., /tmp/music.mp3).

Pro model: Replace lyria-3-clip-preview (30s clips) with lyria-3-pro-preview (full songs with lyrics).

---

6. Multi-turn Image Editing

Edit an existing image through conversation. Uses a temp file for the request body because base64 image data exceeds shell argument limits.

Step 1: Build request JSON

python3 -c "
import json,base64
with open('INPUT_IMAGE_PATH','rb') as f:
    b64=base64.b64encode(f.read()).decode()
payload=json.dumps({
    'contents':[{
        'role':'user',
        'parts':[
            {'inlineData':{'mimeType':'image/png','data':b64}},
            {'text':'EDIT_INSTRUCTION'}
        ]
    }],
    'generationConfig':{'responseModalities':['TEXT','IMAGE']}
})
with open('/tmp/gemini_edit_req.json','w') as f:
    f.write(payload)
"

Step 2: Send request and save result

curl -s -X POST "https://generativelanguage.googleapis.com/v1beta/models/gemini-2.5-flash-image:generateContent?key=$GEMINI_API_KEY" \
  -H "Content-Type: application/json" \
  -d @/tmp/gemini_edit_req.json | python3 -c "
import sys,json,base64
d=json.load(sys.stdin)
if 'error' in d:
    print(f'ERROR: {d[\"error\"][\"message\"]}'); sys.exit(1)
for p in d['candidates'][0]['content']['parts']:
    if 'inlineData' in p:
        img=base64.b64decode(p['inlineData']['data'])
        path='OUTPUT_PATH'
        with open(path,'wb') as f: f.write(img)
        print(f'Saved: {path} ({len(img)} bytes)')
    elif 'text' in p and p['text'].strip():
        print(p['text'])
"

Replace: INPUT_IMAGE_PATH with the source image, EDIT_INSTRUCTION with the edit command, OUTPUT_PATH with the output file path.

For multi-turn editing, include previous turns in the contents array when building the JSON:

payload = json.dumps({
    "contents": [
        {"role":"user", "parts":[{"inlineData":{"mimeType":"image/png","data":"..."}}, {"text":"Draw a red circle"}]},
        {"role":"model", "parts":[{"inlineData":{"mimeType":"image/png","data":"..."}}]},
        {"role":"user", "parts":[{"text":"Now make it blue"}]}
    ],
    "generationConfig":{"responseModalities":["TEXT","IMAGE"]}
})

---

Error Handling

Error Code	Meaning	Action
429	Rate limit exceeded	Wait and retry. Check tier at aistudio.google.com/rate-limit
403	API key invalid or billing not enabled	Verify key and billing status
404	Model not found	Check model ID spelling
400	Invalid request	Check request body format

Important Notes

• All generation models require Paid Tier 1+ billing on the API key
• Free tier keys show RPM 0/0 for all generation models
• Check your quota at https://aistudio.google.com/rate-limit
• Generated files include SynthID watermarks (invisible)
• Veo video generation is asynchronous and may take 30-120 seconds
• All other generation types are synchronous (5-30 seconds typical)