Native media re-framing · Phase 1
Your media,
re-framed natively.
Reframe makes images bigger and sharper, turns a video link into 4K, and pulls a depth map out of a photo. It is headless (no window, no app — a command you run, or a private service you send a link to). Every line of it is Rust. No Python, no bundled machine-learning runtime.
not deployed · Phase 1 CPU lane first no pricing
What it does
Three jobs, one tool.
Each job re-frames your media: a new size, a new resolution, or a new depth channel.
Re-scale
Make an image two or four times larger. Classic filters first, then super-resolution models that sharpen as they enlarge. They restore detail; they do not invent new painted detail.
○ planned building now
4K video from a link
Send a link to a video. Reframe fetches it, enlarges every frame, and re-encodes it as 4K with the original audio. A long clip takes hours, so the job survives a restart and picks up where it stopped.
○ planned
Depth
Turn a photo or a clip into a depth map: how far away each pixel is. The first
model is licensed for research only, so every depth result is labelled
research_only and can never ship in a product.
○ planned research only
Status
Honest about what runs today.
Nothing is served yet. Phase 1 builds the slow, exact reference lane on the CPU; every faster lane must match it before it takes a job.
| Lane | What it is | Status |
|---|---|---|
| CPU reference | the exact, slower lane every other lane is checked against | ○ planned building now |
| GPU | Metal on Apple silicon, Vulkan or DirectX elsewhere | ○ planned |
| Video | link in, 4K out, audio kept, resumable | ○ planned |
| Private service | a job queue on our own machines, behind a login | ○ planned |
| Depth | depth maps, research only | ○ planned |
Speed and quality figures: measured pending. They will be shown with the date and the machine they were measured on, never typed by hand.
How it works
One job, in order.
This is the path a video job will take. An image job skips the video steps.
- You send a link. Links only — no uploads. A link keeps big files off the front door.
- Reframe fetches it, carefully. Only secure links, only known media hosts, a size cap and a time limit on every download. Addresses inside our own network are refused.
- The file is opened in a walled-off helper. Decoding runs as a separate, time-limited process. A broken or hostile file fails the job, never the service.
- Each frame is cut into tiles. Tiles overlap a little and are blended back together, so no seams show.
- Each tile is enlarged. The CPU reference lane runs first. A faster lane takes over only after its first tile matches the reference; if it does not, the job stays on the reference lane.
- Frames are put back together. Re-encoded as 4K, original audio copied across, timing kept.
- You fetch your result. It is private to you and expires after a set time.
A long job saves its progress as it goes. If the machine restarts, the job resumes — it never starts over.
Two ways in
A command, or a private service.
Command line
One program, a few commands: reframe upscale,
reframe video, reframe depth. Progress prints as one
line per update, so scripts and agents can read it.
○ planned upscale first
Private service
reframe serve runs on our own machines behind
Zero Trust
access. You post a job, check on it, and fetch the result. It is a tool for our
own work, not a public service.
○ planned
Models and licenses
The license follows each model file.
Every model file has its own license row. No model files are stored in the source code.
| Model | Job | License | Status |
|---|---|---|---|
| Lanczos, bilinear, nearest | re-scale (classic filters) | no model file | ○ planned |
| SPAN | re-scale | checked per file (CC BY 4.0 where published) | ○ planned |
| RealPLKSR | re-scale | checked per file (CC BY 4.0 where published) | ○ planned |
| Depth Pro 2 (Apple) | depth | research only, non-commercial | ○ planned |
| Depth Anything V2 Small | depth (commercial candidate) | Apache-2.0, to be checked | ○ planned |
Depth Pro 2 is research only. Apple licenses its weights for
non-commercial research. Running it on our own machines does not change that. So
every depth result carries a research_only label, and any path that
publishes or ships files refuses a labelled file. Depth stays an evaluation tool
until a commercially licensed model joins the list.
Model files load only in the
safetensors
format. Formats that can run code when opened (.pth, .pt,
.ckpt, .bin) are refused by name. Each file is checked
against a hash when downloaded and again before every use.
Trust
What you can count on.
Who runs it
The riir.ai maintainers, on our own machines. It is an internal tool first.
Your files
Links in, no uploads. Results are private to the caller who made the job and are deleted when they expire. Nothing is re-hosted or published.
Rights
Send only media you own or have the rights to. Third-party video is never re-published.
Price
None. Reframe is not for sale. If that ever changes, pricing is published here first.
Glossary
Words on this page.
- headless
- no window or app; you run a command or call a service
- super-resolution
- a small model that enlarges an image and adds back real-looking detail
- tile
- a small square piece of a frame, processed on its own and blended back
- reference lane
- the exact CPU version every faster version is checked against
- depth map
- an image where each pixel says how far away that point is
- safetensors
- a model file format that holds numbers only and cannot run code
- Zero Trust
- every request proves who it is before it reaches the service