Marigold V2 — pixel-level depth & normals
Converts a single image into high-resolution depth, surface-normal, and albedo maps at pixel level — sharper than MoGe-3 and InfiniDepth. Runs locally (17–29 GB VRAM).
Video Generation Lab
Experimental interface for generating video content using various AI providers. Rendered results can be saved to your permanent cloud storage.
Generation Results
No renders yet — describe your vision above
Select providers and hit Generate to start
Recent Generations
Loading history...
Modern Research
World Sculpt — scene → editable 3D objects
Turns photos/video into a 3D scene split into separate, individually editable objects (move/resize), and can convert Gaussian-splat worlds into real meshes.
Fire 3D — simulation-ready scenes in under a minute
Reconstructs a simulation-ready 3D scene with separate meshes + materials from photos/video, processing up to 16 objects in parallel on one GPU.
Unimate — animate any 3D skeleton
One model animates arbitrary 3D characters (birds, snakes, dragons, satellites) from a rigged model + a text instruction — no per-character retraining.
The roundup clip was the lead; the project page is the paper. UniMate (SIGGRAPH Asia 2026, Princeton/Berkeley/MIT) is one unified model for motion on arbitrary skeletal topologies — its own claim is no test-time optimization and no per-skeleton retraining — trained on UniML3D, 13,006 motion sequences across bipedal, quadrupedal, avian, marine, insectoid and serpentine rigs, and it supports zero-shot editing, in-betweening and expansion. Checked 2026-10-07.
Roundup this came from — AI News roundup ↗
NVIDIA Soul Refiner — one-step 4K video upscaling, with a fidelity cost
Upscales low-resolution video to 4K in a single step and is model-agnostic (Minimax H3, Cosmos, Wan). Sharpness is a clear gain; fidelity is not — the reviewer's own examples show barn and scene details changing, so treat it as a re-render rather than a faithful enlargement. Open code, fine-tuned from LTX Refiner.