Meridian (Viggle AI)
Geometry-guided video re-rendering from viggle-ai, released 2026 on minimax-h3's transformer and VAE. Read from the model card and its inference docs; arrived by Telegram 2026-09-16. No paper.
What it does
Take an existing clip and revisit it from a new camera: orbit, zoom, pan, rise, with time handled separately — freeze a frame, slow or speed the action — or start from a single image and add a camera move. The pipeline is explicit on the card: Meta's VGGT-Omega estimates depth and camera poses, a 3D point cloud is rendered along the requested camera path, and the H3 transformer fills the holes and refines the result. The text encoder is dropped at inference; precomputed task embeddings stand in for it.
Shape and cost
- 33B parameters; output 24 fps at 1344×768, 73–243 frames (3–10 s).
- Ships as a 2.5 GiB LoRA/adapter by default or a 61.7 GiB full bf16 checkpoint.
- On one B200 with the adapter: 73 frames in ~36 s at 88 GiB peak; 124 frames ~80 s / 89 GiB; 243 frames ~150 s / 113 GiB — the long setting does not fit a single 80 GB card.
- Limits the card states itself: unseen surfaces are generated, not recovered; large viewpoint changes get unstable; retiming cannot invent motion that was never filmed.
Licences, three of them
Weights under the MiniMax H3 Community License, which excludes the EU, UK, South Korea and US; code Apache 2.0; VGGT-Omega under Meta's non-commercial research licence. So the model is open-weight in the sense of downloadable and closed in the sense that matters to most of this spoke's readers: the geography carve-out is inherited from the base, and the geometry front end is non-commercial regardless.
Where it sits
Third page on the minimax-h3 base after studio-1939-lora — but where the style LoRA changed what the frames look like, this changes where the camera is, with an external 3D estimator doing the work the diffusion model cannot. It is the fourth non-Google image/video-generation source in this spoke (mage-flow-turbo, studio-1939-lora, minimax-h3, now this), past the ≥3 spin-out trigger the synthesis has carried since 2026-07-22 and recorded as fired on 2026-09-10.
Tier T3 — first-party card; the throughput numbers are the vendor's on the vendor's hardware, and there is no quality benchmark at all.