One picture in.
A 3D place out.

One command. Two ways in.

Start from a picture you already have:

assetyard generate environment --reference-image assets/hall.png --out assets/hall.glb

Or give only a prompt. The server first paints a landscape reference picture, and then rebuilds it:

assetyard generate environment --prompt "A long stone dungeon hall with a flat flagstone floor" --out assets/hall.glb

assetyard generate model --provider environment is the same request, so a script that already uses it keeps working:

assetyard generate model --provider environment --reference-image assets/hall.png --out assets/hall.glb

The rebuild costs 0 credits during the free beta, the same as mesh work from the 3D modeler. A request with --reference-image costs 0 credits. A request with only a prompt pays the image price for the reference picture that the server paints. The rebuild takes seconds of GPU time. A prompt adds the time to paint the picture.

Add --seed N to set the seed. The flags that every generate command has work here too: --wait, --out, --max-credits, --request-id, --collection, --json.

An environment request rejects the settings that shape an object: a geometry or orientation profile other than source, a face count, a texture resolution, a rig, animation packs or clips, and a body plan.

What is in the GLB

One file holds what a game needs to stand in the picture. The axes follow glTF: +Y is up, -Z is forward, and one unit is one metre. When the backend finds a floor, the floor lies on y = 0 and the camera stands above the origin.

The build read this excerpt from environment-hall.glb. That file has the nodes backdrop, floor_collider, source_camera. The full record also holds the parameters of the floor fit, and a cells array of 57 rows. In a row, . is walkable, # is blocked and ? is unknown.

{
  "schema_version": 1,
  "units": "meters",
  "frame": "floor_aligned",
  "camera": {
    "fov_x_deg": 71.798,
    "fov_y_deg": 43.985,
    "aspect_ratio": 1.792,
    "image_width": 1536,
    "image_height": 857
  },
  "source_view": {
    "node": "source_camera",
    "position": [0, 1.921, 0]
  },
  "floor": {
    "normal": [0, 1, 0],
    "camera_height_m": 1.921,
    "inlier_area_m2": 30.205
  },
  "walkable": true,
  "walkable_grid": {
    "origin": [-1.851, 0, -17.606],
    "columns": 15,
    "rows": 57,
    "cell_size_m": 0.25,
    "walkable_cells": 455
  },
  "collider": {
    "node": "floor_collider",
    "triangles": 40
  }
}

Measured numbers are rounded to 3 decimals. Row 0 of the grid is at the minimum z, and column 0 is at the minimum x.

Use it in a game

The 3D demo plays in these three rooms. The game camera starts at the source view. It pans and zooms to follow the fight, drifts a few centimetres, leans toward the action and pushes in about 36 cm when a fight starts, so the room shows its depth. The game draws the textured backdrop, so a figure passes behind a pillar. A plane far behind the room fills the gaps: the sharp picture where the backdrop has no surface, such as the sky, and a soft, dark copy in the thin cracks at depth jumps. On these rooms, a push-in of half a metre and a sideways move of about 0.2 m stay clean. The game places people and props by a position in the picture, cast from the source camera onto the floor. The walkable grid limits where every character can walk.

curl -O https://assetyard.dev/demo3d/world3d.js

world3d.js is the template. It is one file with no build step. The page writes window.HOLLOWREACH_3D, which names the room files. ROOM_LAYOUT in world3d.js holds the picture positions of the spawn, the exits, the people, the props and the lamps for each room. A new room needs an entry in ROOM_ORDER and in ROOM_LAYOUT.

Write a picture that rebuilds well

• A wide view from slightly above. A raised viewpoint keeps a lot of floor in frame.
• A broad flat floor in the lower half. The walkable grid can only cover floor that the picture shows.
• Even light. The picture is the texture, so its light stays the light of the room.
• No people and no text. A person becomes part of the backdrop, and a game cannot move them.

The three rooms used these prompts. Each one names the view, the floor and what to leave out.

The Last Gate

Elevated wide view from an upper window over a small walled frontier town square at dusk. A broad open square of wet cobblestones fills the lower half of the frame. Market stalls with awnings line the left and right edges. Iron lanterns on wooden posts glow warm amber. A tall stone gatehouse with a closed iron portcullis stands in the far wall. Crenellated town walls, a violet dusk sky. Painterly realistic fantasy game environment, sharp detail, even depth of field. No people, no characters, no animals, no text, no logo, no watermark.

assetyard generate image \
  --prompt "Elevated wide view from an upper window over a small walled frontier town square at dusk. A broad open square of wet cobblestones fills the lower half of the frame. Market stalls with awnings line the left and right edges. Iron lanterns on wooden posts glow warm amber. A tall stone gatehouse with a closed iron portcullis stands in the far wall. Crenellated town walls, a violet dusk sky. Painterly realistic fantasy game environment, sharp detail, even depth of field. No people, no characters, no animals, no text, no logo, no watermark." \
  --width 1664 \
  --height 928 \
  --seed 32 \
  --out assets/environment-town.png

The Long Hall

Elevated wide view from high in one corner of a long vaulted stone dungeon hall. A broad flat flagstone floor fills the lower half of the frame. Thick square stone pillars stand along both sides. Iron wall sconces hold low orange flames. An arched doorway glows at the far end. Thin cold mist lies on the floor. Painterly realistic fantasy game environment, sharp detail, even depth of field. No people, no characters, no skeletons, no text, no logo, no watermark.

assetyard generate image \
  --prompt "Elevated wide view from high in one corner of a long vaulted stone dungeon hall. A broad flat flagstone floor fills the lower half of the frame. Thick square stone pillars stand along both sides. Iron wall sconces hold low orange flames. An arched doorway glows at the far end. Thin cold mist lies on the floor. Painterly realistic fantasy game environment, sharp detail, even depth of field. No people, no characters, no skeletons, no text, no logo, no watermark." \
  --width 1664 \
  --height 928 \
  --seed 32 \
  --out assets/environment-hall.png

The Warden's Vault

Elevated wide view from a stone gallery into a vast circular underground vault. A broad flat cracked stone floor fills the lower half of the frame. A round iron seal in the centre of the floor glows amber from below. Heavy chains hang from the dark domed ceiling. Arched alcoves ring the curved walls. Painterly realistic fantasy game environment, sharp detail, even depth of field. No people, no characters, no statues, no text, no logo, no watermark.

assetyard generate image \
  --prompt "Elevated wide view from a stone gallery into a vast circular underground vault. A broad flat cracked stone floor fills the lower half of the frame. A round iron seal in the centre of the floor glows amber from below. Heavy chains hang from the dark domed ceiling. Arched alcoves ring the curved walls. Painterly realistic fantasy game environment, sharp detail, even depth of field. No people, no characters, no statues, no text, no logo, no watermark." \
  --width 1664 \
  --height 928 \
  --seed 32 \
  --out assets/environment-vault.png

What one picture cannot do

• One picture shows one side of a place. The result looks right near the viewpoint of the picture. It has nothing behind objects.
• The camera stays near the viewpoint. A game camera can pan, zoom and turn a little. Characters can walk anywhere on the walkable floor.
• The floor must be visible. A surface tilted more than 45 degrees from level is not a floor.
• Metric scale is an estimate. Check a size in metres against something you know, such as a door.
• No floor means no collider. A picture without a visible floor still returns a backdrop, but no floor collider and no walkable grid.

License

MoGe-2 comes from Microsoft. Its code and weights use the MIT license. Its DINOv2 backbone uses Apache-2.0. The three rooms are part of the free pack, and the pack terms apply to them.