Cloud Instances

Cloud Instances

Recaster uses Vast.ai to provision cloud GPU instances for remote training, Quick Recast processing, and video upscaling. Each instance type is optimized for its specific workload.

StudioCloud instances require a Studio tier license.

Instance Types

Recaster supports three instance types, each configured and provisioned for a specific workload. When you launch an instance, you select the type and Recaster handles the rest — from GPU selection to software installation.

TypeDiskProvisioningUse Case
Training40+ GB3 – 5 minSAEHD, AMP, Quick96, and XSeg model training with live preview
Quick Recast20 GB2 – 3 minFace swapping and enhancement with advanced models (Ghost, CodeFormer)
Upscale20 GB2 – 3 minAI video upscaling with Real-ESRGAN, SwinIR, and full pipeline support

Launching an Instance

1

Open the Remote panel

Click the Remote tab in the sidebar, or enable Remote mode from the Project Panel toggle switch.
2

Select instance type

Choose Training, Quick Recast, or Upscale from the instance launcher. Each type shows the estimated disk requirement and provisioning time.
3

Choose a GPU

Recaster queries Vast.ai for available GPU offers and displays them sorted by price. You can filter by GPU model, VRAM, and price per hour. Select the GPU that fits your budget and performance needs.
4

Launch and wait

Click Launch to begin provisioning. The instance boots, installs the required software stack, and creates the workspace directory structure. A progress indicator shows each provisioning stage.
5

Instance ready

Once provisioning completes, the instance card in the Remote panel shows a green status indicator with the instance type badge. You can now use it for processing.

GPU Selection Guide

The GPU you choose affects both processing speed and cost. Here are guidelines for common scenarios:

Budget Option

  • RTX 3070 / 3080 (8 – 10 GB VRAM)
  • $0.10 – $0.30 per hour
  • Good for Quick Recast and upscaling
  • Adequate for training shorter sessions

Best Value

  • RTX 3090 / 4080 (16 – 24 GB VRAM)
  • $0.30 – $0.70 per hour
  • Excellent for all workloads
  • Recommended for training and upscaling

Maximum Performance

  • RTX 4090 / A100 (24 – 80 GB VRAM)
  • $0.70 – $1.50 per hour
  • Fastest processing across all tasks
  • Best for long training sessions and 8x upscaling

Avoid

  • GPUs with less than 8 GB VRAM
  • Very old architectures (pre-Turing)
  • Instances with high disk-to-GPU ratios
  • Offers significantly above market rates

Check pricing regularly

Vast.ai GPU pricing fluctuates based on supply and demand. The same GPU model can vary significantly in price between hosts. Sort by price and check a few options before launching.

Provisioning Process

When an instance launches, Recaster runs an automated provisioning script that installs all required dependencies. The provisioning process is specific to each instance type:

Training Instance

Installs DeepFaceLab, Python dependencies, and the streaming preview server. Sets up the workspace with source and destination face directories. Configures the X11 environment for headless training with preview capture.

Quick Recast Instance

Installs ONNX Runtime GPU 1.19.2, OpenCV, cuDNN 9, and the Quick Recast server script. Creates workspace directories for source, target, and output files. Models are downloaded on demand during first use.

Upscale Instance

Installs ONNX Runtime GPU 1.19.2, NumPy (<2.0), OpenCV, and cuDNN 9. Creates workspace at /workspace/upscale/ with input, output, models, and previews subdirectories. Verifies CUDA provider availability.

SSH Keys

Recaster automatically generates and manages SSH keys for connecting to cloud instances. Keys are stored in the application data folder:

  • macOS: ~/Library/Application Support/Recaster/ssh/
  • Windows: %APPDATA%\Recaster\ssh\
  • Linux: ~/.config/Recaster/ssh/

Keys are generated the first time you create an instance and are reused for subsequent instances. The public key is registered with Vast.ai, and the private key stays on your machine.

Key permissions

SSH keys must have restricted permissions (chmod 600 on macOS and Linux). Recaster sets this automatically, but if you encounter SSH connection errors, verify the key file permissions.

Instance Lifecycle

Cloud instances go through the following states during their lifecycle:

  1. Launching — Vast.ai is allocating the GPU and booting the instance.
  2. Provisioning — Recaster is installing dependencies and setting up the workspace.
  3. Running — The instance is ready for use. This is when billing is active.
  4. Stopping — The instance is shutting down. No new tasks can be started.
  5. Stopped — The instance is no longer running. No charges are incurred. Data on disk may or may not persist depending on the Vast.ai host.
  6. Deleted — The instance and all associated data are permanently removed.

Stop instances when not in use

Cloud GPU instances are billed per hour while running, even if no processing is active. Always stop or delete instances from the Remote panel when you are finished. See Cost Tracking to set up budget alerts.

Project Association

Each Recaster project can be associated with a cloud instance. When you toggle Remote mode in the Project Panel, the project remembers the associated instance ID. On subsequent launches, if the same instance is still running, Recaster automatically reconnects.

If multiple instances are running when you enable Remote mode, a selection dialog lets you choose which instance to associate with the current project. If only one instance is available, it is selected automatically.

File Synchronization

The Sync tab in the Project Panel provides file synchronization between your local project and the remote instance. You can upload project files (source faces, training data) to the instance and download results (trained models, processed videos) back to your local machine.

  • File transfer uses rsync over SSH for efficient incremental sync
  • Progress bars show real-time transfer status
  • Color-coded status indicators show which files are synced, local-only, or remote-only
  • Bulk upload and download operations for entire directories

Troubleshooting

Instance fails to provision

Some Vast.ai hosts may have configuration issues. Try deleting the instance and launching on a different host. Check that the selected GPU has sufficient disk space for the instance type.

SSH connection refused

Wait 1 to 2 minutes after the instance shows as "Running" before trying to connect. The SSH server may still be starting up. If the issue persists, check that your SSH key permissions are correct (chmod 600).

Remote mode toggle not visible

The Remote mode toggle only appears for Studio tier users. Check your license status in the status bar at the bottom-right of the main window. If it shows "Free", activate your Studio license via Help > Activate License.

Was this page helpful?