FrameLatheDocs

Hardware

Local GPU on Windows. Remaster needs NVIDIA CUDA. Other presets can fall back to ncnn or CPU.

Enhancement runs on this computer. There is no cloud GPU to burst to.

Windows

The shipping installer is Windows. macOS is not produced or tested.

GPU paths

  • Export Ready, Denoise, Smooth 60 fps can use the GPU in the machine. On NVIDIA they prefer the CUDA stack when the GPU pack is installed. Without NVIDIA, ncnn or CPU paths may still run upscale and denoise.
  • Remaster requires NVIDIA CUDA, the GPU add-on pack when you request it, VSR weights, and roughly 12 GB VRAM. See Remaster.
  • Clean 8K is an advanced Export-style upscale to 4320. It is extremely slow on mid-range cards.
  • The photo check runs in the app window through WebGPU. On a PC with two graphics chips, FrameLathe asks Windows for the faster one. Without WebGPU it runs on the processor, more slowly.
  • The face fix also runs in the app window through WebGPU, and needs a graphics card that reports half-precision (shader-f16) support and lets one buffer hold at least 1 GB. On a PC without one it is not offered, and the panel says so in one line; Sharpen faces still runs.

The hardware strip

The strip at the top of the app names your graphics card and its memory. Hover any part of it for what it means:

  • The card name says what this GPU will run and whether Remaster can run on it (Remaster needs NVIDIA CUDA and roughly 12 GB VRAM).
  • Fastest path ready means a TensorRT engine for this GPU is already built and upscale reuses it.
  • Faster path not built yet means upscale works now, on the slower route. Build it once opens Settings → Advanced → Hardware to build a TensorRT engine for this GPU; it is built once and reused for later jobs.
  • CUDA ready and Vulkan path say which route upscale takes today. Set up CUDA opens Settings → Advanced → Hardware for the one-time CUDA install.
  • VRAM … GB free is graphics memory free right now. Other GPU apps use it too.
  • RAM ⚠ means this PC’s own memory (not the card’s) is nearly full. Close other applications or run fewer jobs at once.

Video tools

The installer does not carry ffmpeg or the upscale and frame-rate tools. The app downloads them from framelathe.com on first launch (about 290 MB, roughly 500 MB on disk) into its data folder and keeps them across updates.

GPU add-on pack

The lean engine sidecar does not include torch. The optional GPU engine is a large download (~2.04 GB) from PyPI / PyTorch indexes, triggered by you, not by opening the app.

It carries TensorRT for RTX. A pack downloaded before TensorRT for RTX joined it does not, so Settings → Advanced → Hardware shows TensorRT for RTX as not installed while CUDA is ready, with a Re-download GPU engine button that replaces the pack.

What the app will not do

  • It will not upload a clip to someone else’s datacenter because your GPU is busy.
  • It will not silently swap Remaster for a cloud API.
  • It may degrade denoise to FFmpeg hqdn3d when the neural denoise backend is not ready, and the row says so when it does.