Models

Model Hub is a registry of stage slots. Upscale, Denoise, and Matting take your own weight. Colorize stays first-party.

FrameLathe can fetch model weights you choose. That is optional network use, not cloud rendering.

Model Hub

Opening Model Hub loads the public model index (OpenModelDB, for upscalers). That request does not include your filenames or pixels.

The Hub is a registry of stage slots, not one custom-model flag. Three of those slots accept a file; Colorize runs first-party only.

It is one list. Search it by name, scale, architecture or licence; switch between All, Upscale, Denoise and Matting; and Installed only shows what is on this machine. The count beside the filters says how many models are shown. Import a file… sits at the foot of the list.

Each row has one action: Download, a progress bar while it downloads, or Installed. The arrow at the end of a row opens its details: the id, architecture and licence, a link to its OpenModelDB page, and for an installed model Use for Upscale (or Denoise, Matting), which applies that weight to the current recipe, and Remove. Download is not the same as Use — an installed model is not used until you pick it, and its row then reads Installed · in use. Queue chips read Upscale · name, Denoise · name, Matte · name.

The recipe rail Models block shows a picker only for slots that recipe actually runs. Browse Hub opens the Hub on that stage.

A bring-your-own weight is a choice, not automatically an upgrade — FrameLathe does not publish scores for Hub models. Colorize is not swappable.

Photos

Photos use the same first-party models as video, and none of them is swappable from the Hub:

Copies FrameLathe hosts

Three of these come from framelathe.com rather than their publishers’ servers. Each is licensed to you by its publisher, not by FrameLathe, and its folder on your computer gets the publisher’s licence (LICENSE) and a NOTICE saying what FrameLathe changed:

• DDColor (piddnad/DDColor-models, Apache-2.0): converted to ONNX by FrameLathe from the published PyTorch weights, at full precision. No retraining.
• SigLIP 2 base (google/siglip2-base-patch16-224, Apache-2.0): the vision tower only, converted to ONNX and quantised to 4 bits by FrameLathe. No retraining.
• Stable Diffusion 1.5 inpainting (stable-diffusion-v1-5/stable-diffusion-inpainting, CreativeML OpenRAIL-M): converted to ONNX and to mixed 16-bit precision by FrameLathe, with the text encoder replaced by the fixed embeddings of one prompt. No retraining. The licence’s use restrictions apply to you, and the ONNX files carry the licence and notice too.

No safety checker. Stable Diffusion’s optional safety checker is not included. The licence does not require it, the face fix cannot be prompted (fixed embeddings stand in for the text encoder), and it only redraws an area of your own scan that you mark.

About Stable Diffusion 1.5. The face fix uses Stable Diffusion 1.5 inpainting, an openly licensed model trained on large web-image datasets that have been criticised for what they contained. We chose it because it does this job well and runs on your own computer. It is off by default, it downloads only after you accept its licence, it cannot be prompted (it only redraws the area of your own scan that you mark), and every face it draws is invented, not recovered. In August we replaced an earlier fill-light feature built on this model with a classical method. We will replace it when a suitable alternative exists.

GPU pack vs models

• GPU add-on pack — NVIDIA / PyTorch wheels so CUDA jobs can run. Upscale import and OpenModelDB download need this pack on a packaged install.
• Model weights — networks the engine runs locally after the file is on disk.

Both happen because you asked. Neither sends your video to FrameLathe.

Hub licence filter

A filter is on by default. While it is on, FrameLathe will not download, import, plan, or start a Hub or imported weight whose recorded licence string looks non-commercial, research-only, academic-only, or is missing / unrecognised. The label comes from OpenModelDB or the file you imported, not from FrameLathe reading the publisher’s contract. The filter is the Commercial use OK switch beside the search. While it is on, blocked models are left out of the list; a model you already installed stays listed, so it can still be found and removed.

The first time you open the Model Hub, it shows this notice in place of the list, with Keep the filter on and Turn the filter off…. Nothing can be downloaded or imported from the Hub until you choose one.

Turn it off in Settings → Advanced → Model Hub if you supplied the file and accept responsibility for its publisher’s terms. Turning it off does not grant rights. Turning it on does not mean remaining models are cleared for your use.

This filter only applies to Upscale, Denoise, and Matting — the three slots that take a file. Turning it off changes what those slots accept; it does not open a Colorize slot, because there is none, so a colorizer ONNX has nowhere to go.

Imported denoise and matte ONNX files record license: unknown, so the filter blocks them until you turn it off.

First-party presets and the weights listed under Settings → Help & about → Legal are not gated by this filter.

Contracts

Denoise: one 4D 3-channel input; one 3-channel output at the same height and width; float32; RGB 0–1. Dynamic spatial dims are preferred. A static graph is tiled or resized.

Matting: one 4D 3-channel RGB input; one 1-channel alpha at the same spatial size (resized back to the frame if the graph is static); float32; 0–1. Generic RGB→alpha. There is no fill or inpaint slot.

Upscale stays on spandrel image SR. Restoration and 1× weights cannot stand in as the upscaler.

Third-party terms

Hub models keep their own licences. The EULA does not promise commercial-use rights for every weight. Read Settings → Help & about → Legal on the installed app.