Lip-sync batches on a $0.27/h RTX 3090 with LatentSync 1.6

Description and use

Guide to batch lip‑sync on a rented RTX 3090 at $0.27/h using LatentSync 1.6. Covers offer selection, environment setup, queue organization, and cost calculation.

Problem

Align existing talking‑head video with speech audio at low cost using rented GPU resources.

Outcome

A single 8‑second clip was processed in 238 s on a rented RTX 3090, costing ~$0.09. Setup took ~9 min. The queue system allowed inspection of failures and ensured GPU utilization.

Workflow

• 1. Search and select a Vast.ai offer with RTX 3090, reliability ≥0.97, free inbound traffic, ≥150 GB disk, compatible CUDA driver.
• 2. Create instance with chosen offer, using pytorch/pytorch:2.5.1‑cuda12.4‑cudnn9‑runtime image.
• 3. Install build‑essential and LatentSync dependencies, download 5 GB weights from Hugging Face.
• 4. Prepare input video and audio locally before renting.
• 5. Run a known test job to verify inference command and output.
• 6. Create a queue adapter script that accepts video, audio, and output paths.
• 7. Organize files in queue/, done/, output/, failed/ directories.
• 8. Execute run_queue.py to process jobs sequentially, moving successful pairs to done/ and failures to failed/.
• 9. Download results before terminating instance to avoid storage billing.
• 10. Destroy instance after batch completion.

Evidence

We ran LatentSync 1.6 on a rented RTX 3090 at $0.27/h.

Source

Our 720×1280 talking-head clip with eight seconds of TTS audio took 238 seconds to process, using the 512-pixel stage, 20 inference steps, guidance scale 1.5, and DeepCache.

Source

Setup took roughly nine minutes.

Source

The recorded clip cost includes more than the 238 seconds spent processing.

Source