https://www.eclair.earth/
Image-to-Video Generation Subnet with GPT-4o Evaluation
Miners run image-to-video (I2V) generators. Validators sample real video clips, extract the first frame, generate a description via GPT-4o, and challenge miners to generate video from the frame + prompt. Scoring uses GPT-4o forced-choice comparison: which video looks more real? Winner takes all weights.
┌─────────────────────────────────────────────────────────────────────────┐
│ SAMPLE GENERATION LOOP │
├─────────────────────────────────────────────────────────────────────────┤
│ 1. Download random video from Hippius "lot-of-videos" bucket │
│ 2. Extract 5s clip and first frame │
│ 3. Generate prompt via GPT-4o description │
│ 4. Query each miner's I2V model │
│ 5. Score each miner vs original using GPT-4o forced choice │
│ 6. Upload sample to "video-samples" bucket with full metadata │
└─────────────────────────────────────────────────────────────────────────┘
┌─────────────────────────────────────────────────────────────────────────┐
│ WEIGHT SETTING LOOP │
├─────────────────────────────────────────────────────────────────────────┤
│ 1. Read all samples from "video-samples" bucket │
│ 2. Calculate win rates per miner (wins / total) │
│ 3. Apply winner-take-all with epsilon beat rule │
│ 4. Set weights on chain (leader gets 1.0, others get 0.0) │
└─────────────────────────────────────────────────────────────────────────┘
# 1. Clone the repo
git clone https://github.com/your-org/eclair.git
cd eclair
# 2. Create .env from example
cp env.example .env
# Edit .env with your API keys and wallet info
# 3. Run
docker compose up -d# Requires Python 3.12+, ffmpeg
pip install uv
uv pip install -e .
# Run
eclair| Variable | Required | Description |
|---|---|---|
CHUTES_API_KEY |
Yes | Chutes API key for I2V generation |
OPENAI_API_KEY |
Yes | OpenAI API key for GPT-4o evaluation |
HIPPIUS_SEED_PHRASE |
Yes | Hippius subaccount seed phrase |
WALLET_NAME |
No | Bittensor wallet name (default: default) |
HOTKEY_NAME |
No | Bittensor hotkey name (default: default) |
NETWORK |
No | Bittensor network (default: finney) |
Miners commit their I2V chute slug to chain. The validator will call your chute with:
{
"prompt": "A person walks through a park...",
"image": "<base64 encoded first frame>",
"fps": 16,
"frames": 81,
"resolution": "480p",
"fast": true
}Your chute should return the generated video as video/mp4.
Commit JSON to chain:
{"generator_chute": "your-chute-slug"}- Metric: GPT-4o forced choice - which video (original vs generated) looks more real?
- Win: Generated video chosen as more real than the original clip
- Score: Win rate = wins / total samples
- Weights: Winner-take-all (leader must beat all predecessors by epsilon)
All samples are stored in Hippius S3 for reproducibility:
s3://video-samples/
└── 2024-01-29_12-34-56/
├── original_clip.mp4
├── first_frame.png
├── miner_5FHneW46.mp4
├── miner_8Abc1234.mp4
└── metadata.json
Each metadata.json contains:
- Source video info (bucket, key, clip timing)
- GPT-4o generated prompt
- Per-miner evaluation results (wins, confidence, reasoning, artifacts)
- All hotkeys and chute slugs
MIT