A quick check of where today's world models are, including ones you can host yourself. Every model got the same single photo. For each, see the video it can produce and the 3D world you can walk through, with what it cost to run. Ranked best first.
Only NVIDIA Lyra generates a real video. The other world models output 3D directly, so they have no video.
Ranked by how convincing the result is, judged by eye on this one photo (9–10 Oct 2026). Lyra total across all runs: $7.29 on the Modal dashboard, from the free $30 monthly credit. Marble used the World Labs API's included credits. Everything else ran free.
The full NVIDIA Lyra story (inputs, intermediate outputs, distilled vs full quality, a refinement experiment) is on its own page. We also tested whether a coding agent (Claude Code writing Three.js) or a hybrid of code and a world model does better than world models alone:
About 1,000 lines of procedural code: a parametric spiral stair, a ribbed glass dome and instanced palms. Works from every angle; you can walk the floor and climb to the top gallery. It looks stylised next to the photo.
Open the coded scene
The world model's own stair is cut out and the coded one takes its place, aligned to the photographer's position. Photographic palms and light, plus a stair you can climb and edit.
Open the hybridJudged by eye in the viewers on 9 October 2026; no automated metrics yet.
| What we checked | Coding agent | SHARP | HY-World 2.0 | Lyra 2.0 | Hybrid |
|---|---|---|---|---|---|
| Looks like the photo (from its viewpoint) | Recognisable, stylised | Near-photographic | Stair sharp, sky and glass missing | Photographic; filigree and roof kept | Photographic plants, stylised stair |
| Holds up when you move | Every angle | Breaks after 1–2 m | Breaks after 1–2 m | Holds along the generated path | Stair yes, plants break |
| Fine ironwork (filigree, balusters) | Approximated by textures | Smoothed away | Pierced pattern kept | Kept | Coded approximation |
| Behind the stair / outside the frame | Whole glasshouse built | Empty | Empty | Invented (orbit run) | Coded stair only |
| Walk and climb | Yes | No | No | No | Yes (stair) |
| Editable (steps, radius, layout) | Change a parameter | No | No | No | Stair only |
| Time to result | One agent session | ~8 min (mostly GPU queue) | ~14 min queue, 0.73 s inference | ~11 min / ~44 min of one H100 (~$0.74 / ~$2.92), distilled / full | Both, plus alignment |
| Download size | 52 KB of code | 16 MB (500k splats) | 8.3 MB (271k splats) | ~60 MB (4M of 11.2M splats, .spz) | ~15 MB |
From the photo's own viewpoint, both splats beat hand-written code on realism, especially the foliage, which is the hardest thing to code.
Single-shot models (SHARP, HY-World 2.0) only know what the camera saw. Lyra generates the unseen views as video first, so its 3D has a believable back side. Code still wins on walking, climbing and editing.
Let code own the structure you interact with and the world model own the texture-heavy surroundings. The two needed only a scale fit to line up: SHARP's metric depth matched the coded stair within 8%.
Prompts for hosted world models that generate whole explorable worlds rather than reconstructing one view. Upload the photo as the image input where the tool accepts one.
Interior of a 19th-century Victorian palm house conservatory in Vienna. At the centre, a white-painted cast-iron spiral staircase winds anticlockwise around a slender central column up to a circular gallery high under the roof. The stair has pierced filigree risers, scrolled side panels, beaded balusters and a curved handrail, with a gothic-arched iron gate at its base. Above, a tall curved glass roof on cream iron ribs, with scalloped overlapping glass panes and a large diagonal tubular iron beam. Dense tropical planting all around: Mediterranean fan palms, a tall dark conifer, ferns and broadleaf climbers. Bright, hazy daylight through the glass, soft shadows, a gravel path around the stair's base.World 1: photo + this prompt. World 2: the same with Expand on, to see what it invents behind the stair. World 3: block out a cylinder (stair) and a dome (roof) in Chisel first.
A walkable Victorian glasshouse. In the middle stands a three-storey white cast-iron spiral staircase with perforated lace-like risers and slender turned balusters, rising to a round viewing gallery just below a soaring curved glass roof. Fan palms, tree ferns and climbing vines fill the space; a gravel path circles the stair. Midday sun diffused through old glass, warm cream ironwork, deep greens, gentle haze. Camera starts at ground level a few metres from the stair, looking up.Useful as a control: how much does the photo itself help compared with a description?
First-person explorer inside a sunlit Victorian palm house. Ahead, a white cast-iron spiral staircase with filigree steps winds up to a gallery under the glass roof. Walk forward along the gravel path, circle the staircase once, then climb it to the top gallery and look out over the palm canopy.Score it on whether the staircase keeps the same shape as you circle it, and whether you can actually climb it.
world-model/run_spaces.py sends the photo to the HY-World 2.0 and SHARP Hugging Face Spaces (a free HF token raises the GPU quota).world-model/ply_to_splat.py converts each .ply to the compact .splat format used here.world-model/cut_stair.py fits each splat's scale to the coded stair and cuts out the world model's own stair for the hybrid.site/coded/js/ holds the coded scene: staircase.js, glasshouse.js, plants.js.