HunyuanVideo-1.5 Text-to-Video Implementation - Complete ✅

Summary

Successfully implemented HunyuanVideo-1.5 text-to-video generation with modular architecture, following separation of concerns principles.

Implementation Details

1. Service Structure ✅

File: backend/services/llm_providers/video_generation/wavespeed_provider.py

HunyuanVideoService: Complete implementation
- Model-specific validation (duration: 5, 8, or 10 seconds, resolution: 480p or 720p)
- Based on official API docs: https://wavespeed.ai/docs/docs-api/wavespeed-ai/hunyuan-video-1.5-text-to-video
- Size format conversion (resolution + aspect_ratio → "width*height")
- Cost calculation ($0.02/s for 480p, $0.04/s for 720p)
- Full API integration (submit → poll → download)
- Progress callback support
- Comprehensive error handling

2. Unified Entry Point Integration ✅

File: backend/services/llm_providers/main_video_generation.py

_generate_text_to_video_wavespeed(): New async function
- Routes to appropriate service based on model
- Handles all parameters
- Returns standardized metadata dict
ai_video_generate(): Updated
- Now supports WaveSpeed text-to-video
- Default model: hunyuan-video-1.5
- Async/await properly handled

3. API Integration ✅

Model: wavespeed-ai/hunyuan-video-1.5/text-to-video

Parameters Supported:

✅ prompt (required)
✅ negative_prompt (optional)
✅ size (auto-calculated from resolution + aspect_ratio)
✅ duration (5, 8, or 10 seconds)
✅ seed (optional, default: -1)

Workflow:

✅ Submit request to WaveSpeed API
✅ Get prediction ID
✅ Poll /api/v3/predictions/{id}/result with progress callbacks
✅ Download video from outputs[0]
✅ Return metadata dict

4. Features ✅

✅ Pre-flight validation: Subscription limits checked before API calls
✅ Usage tracking: Integrated with existing tracking system
✅ Progress callbacks: Real-time progress updates (10% → 20-80% → 90% → 100%)
✅ Error handling: Comprehensive error messages with prediction_id for resume
✅ Cost calculation: Accurate pricing ($0.02/s 480p, $0.04/s 720p)
✅ Metadata return: Full metadata including dimensions, cost, prediction_id

5. Size Format Mapping ✅

Resolution → Size Format:

480p + 16:9 → "832*480" (landscape)
480p + 9:16 → "480*832" (portrait)
720p + 16:9 → "1280*720" (landscape)
720p + 9:16 → "720*1280" (portrait)

6. Validation ✅

HunyuanVideo-1.5 Specific:

Duration: Must be 5, 8, or 10 seconds (per official API docs)
Resolution: Must be 480p or 720p (not 1080p)
Prompt: Required and cannot be empty

Code Structure

backend/services/llm_providers/
├── main_video_generation.py          # Unified entry point
│   ├── ai_video_generate()          # Main function (async)
│   └── _generate_text_to_video_wavespeed()  # WaveSpeed router
│
└── video_generation/                # Modular services
    ├── base.py                       # Base classes
    └── wavespeed_provider.py         # WaveSpeed services
        ├── BaseWaveSpeedTextToVideoService  # Base class
        ├── HunyuanVideoService      # ✅ Implemented
        └── get_wavespeed_text_to_video_service()  # Factory

Usage Example

from services.llm_providers.main_video_generation import ai_video_generate

result = await ai_video_generate(
    prompt="A tiny robot hiking across a kitchen table",
    operation_type="text-to-video",
    provider="wavespeed",
    model="hunyuan-video-1.5",
    duration=5,
    resolution="720p",
    user_id="user123",
    progress_callback=lambda progress, msg: print(f"{progress}%: {msg}")
)

video_bytes = result["video_bytes"]
cost = result["cost"]  # $0.20 for 5s @ 720p

Testing Checklist

Test with valid prompt
Test with 5-second duration
Test with 8-second duration
Test with 10-second duration
Test with 480p resolution
Test with 720p resolution
Test with negative_prompt
Test with seed
Test progress callbacks
Test error handling (invalid duration)
Test error handling (invalid resolution)
Test cost calculation
Test metadata return

Next Steps

✅ HunyuanVideo-1.5: Complete
⏳ LTX-2 Pro: Pending documentation
⏳ LTX-2 Fast: Pending documentation
⏳ LTX-2 Retake: Pending documentation

Notes

Audio support: Not supported by HunyuanVideo-1.5 (ignored with warning)
Prompt expansion: Not supported by HunyuanVideo-1.5 (ignored with warning)
Aspect ratio: Used for size calculation (landscape vs portrait)
Polling interval: 0.5 seconds (as per example code)
Timeout: 10 minutes maximum

Ready for Testing ✅

The implementation is complete and ready for testing. All features are implemented following the modular architecture with separation of concerns.

5.0 KiB Raw Blame History