README
Meridian
A new perspective on space and time.
Change the viewpoint and source time of a video, or move the camera around a single image.
One workspace
- Left — Input + Config: upload a video/image or choose one of two climbing setups, then adjust the camera and time.
- Right — Warp preview + Generated video: inspect the projected source first, then generate your new view.
Click 1. Preview warp → 2. Generate. The warp is a geometric projection, not a generated result; gray areas have no reliable source reference. The generated player stays empty until your generation finishes. There are no galleries, showcase tabs, or pre-generated results in the workspace.
Start with a climbing setup
| Setup | Camera | Source time |
|---|---|---|
| Video · hold + orbit | Orbit left, 30° | Hold at 1.5 s |
| Image · move closer | Move closer, 60% travel | Fixed image |
Both load the input and recorded settings, including seed 42; they do not run inference or load a saved output. These recipes were selected from actual Space generations. Reruns may vary, and changing the camera or time is an experiment—not a guarantee of the same quality. NBA and the 60° climbing orbit are not recommended presets.
Your own input
Use a sharp, continuous video with at least 3.05 seconds remaining after the chosen start, or a single image. Uploads are limited to 100 MB. Orbits target the image center: keep the subject centered and try smaller angles when the warp breaks apart. Large unseen regions and geometry errors can cause artifacts even with a plausible warp.
Hold a moment freezes source time; Let it play uses 1× playback; Slow it down uses 0.5× playback by repeating frames, not interpolating them. Image inputs hide the video-time controls because their source time is fixed. Video start and seed live under More settings.
Changing the source, camera, or time clears the warp and generated video. Changing only the seed clears the generated video but keeps the valid warp. Controls are temporarily locked while a job is queued or running, so its result cannot be mistaken for a different setup. Download the generated video from its player and the recipe with Download settings.
Live inference uses ZeroGPU xlarge. Sign in to Hugging Face with sufficient quota. Each take is 73 frames at 24 fps, using Meridian's teacher with its four-step DMD adapter. The canvas is aspect-dependent; the climbing setups generate at 1024 × 576. Runtime inputs and outputs are temporary—do not upload sensitive footage.
Model and code · Deployment setup
The app code is Apache-2.0. Model and third-party media licenses are separate; the code license does not
relicense the weights or films. See the model repository for weight terms and examples/CREDITS.md for
the sample input clips. VGGT-Omega requires separately approved access under its upstream license.
Contributors
12 commits
Similar repos
README
Meridian
A new perspective on space and time.
Change the viewpoint and source time of a video, or move the camera around a single image.
One workspace
- Left — Input + Config: upload a video/image or choose one of two climbing setups, then adjust the camera and time.
- Right — Warp preview + Generated video: inspect the projected source first, then generate your new view.
Click 1. Preview warp → 2. Generate. The warp is a geometric projection, not a generated result; gray areas have no reliable source reference. The generated player stays empty until your generation finishes. There are no galleries, showcase tabs, or pre-generated results in the workspace.
Start with a climbing setup
| Setup | Camera | Source time |
|---|---|---|
| Video · hold + orbit | Orbit left, 30° | Hold at 1.5 s |
| Image · move closer | Move closer, 60% travel | Fixed image |
Both load the input and recorded settings, including seed 42; they do not run inference or load a saved output. These recipes were selected from actual Space generations. Reruns may vary, and changing the camera or time is an experiment—not a guarantee of the same quality. NBA and the 60° climbing orbit are not recommended presets.
Your own input
Use a sharp, continuous video with at least 3.05 seconds remaining after the chosen start, or a single image. Uploads are limited to 100 MB. Orbits target the image center: keep the subject centered and try smaller angles when the warp breaks apart. Large unseen regions and geometry errors can cause artifacts even with a plausible warp.
Hold a moment freezes source time; Let it play uses 1× playback; Slow it down uses 0.5× playback by repeating frames, not interpolating them. Image inputs hide the video-time controls because their source time is fixed. Video start and seed live under More settings.
Changing the source, camera, or time clears the warp and generated video. Changing only the seed clears the generated video but keeps the valid warp. Controls are temporarily locked while a job is queued or running, so its result cannot be mistaken for a different setup. Download the generated video from its player and the recipe with Download settings.
Live inference uses ZeroGPU xlarge. Sign in to Hugging Face with sufficient quota. Each take is 73 frames at 24 fps, using Meridian's teacher with its four-step DMD adapter. The canvas is aspect-dependent; the climbing setups generate at 1024 × 576. Runtime inputs and outputs are temporary—do not upload sensitive footage.
Model and code · Deployment setup
The app code is Apache-2.0. Model and third-party media licenses are separate; the code license does not
relicense the weights or films. See the model repository for weight terms and examples/CREDITS.md for
the sample input clips. VGGT-Omega requires separately approved access under its upstream license.
Similar repos
Contributors
12 commits