Image to 3D model, prepared for a real game engine

Summer's reference check uses the same rules a good 3D artist would. An image goes straight to conversion only when it shows one complete, isolated subject. For a character that means full body, front or near-front, a neutral symmetric A, T or rest pose, and limbs visibly separated from the torso. For a prop it means a clear three-quarter or product view that reveals the front, a side and the depth of the major forms.

Anything else triggers preparation. Side or rear views, crossed or merged limbs, a subject that is cropped, tiny in frame or mixed with scenery, text or other objects all get one cleanup edit first. Art style does not matter, so pixel art, painted concept art and photos all work if the framing is right.

• Crop to one subject and leave a margin around it.
• Use a plain or removed background. Shadows on the floor often become geometry.
• Avoid a camera directly front-on for props. Three-quarter shows depth the model would otherwise invent.
• For characters, keep hands open and away from the hips so the rig has clean armpits to work with.
• If the back of the design matters, draw or render it and use a multi-view model instead of hoping.
• Keep the file light. Hunyuan 3D Rapid takes a single front-view JPG, PNG or WEBP between 128 and 5000 px and up to 8 MB.

Face counts from image models run high by default. On fal, Hunyuan 3D v3.1 Pro accepts 40,000 to 1,500,000 faces with a default of 500,000. Summer sends a target on every Hunyuan Pro job, at least the 40,000 floor, so a draft does not come back at half a million faces unless you ask for it. Meshy 6 targets about 30,000 faces unless told otherwise, and Rodin rounds your target to its own mesh tiers.

For a hero character you plan to bake, a high count is fine. For props you will scatter, start low, then let the engine help. Godot 4 generates mesh LODs automatically when it imports a glTF scene, so a mid-poly prop already gets cheaper versions for distance. An OBJ imported as a plain mesh gets no LODs and carries no skinning or animation, so GLB is the better format for anything you will rig or place at scale.

Godot recommends glTF 2.0 as its 3D format, and the binary GLB that Studio produces is the version with textures embedded in one file. Summer's download step caps embedded textures at 2048 px, which covers most game use, and keeps node names intact so Godot's import suffixes still work.

With the desktop app open, Studio's import sends the asset to the connected project and runs the normal import pipeline. Rigged character packages land in their own folder under `res://characters/` with the rig GLB, a small character scene and a character.json that lists the clips, scaled to the rig height and facing the engine's forward direction.

From there, ask the agent to place it, snap it to the floor and play the scene. If you build in another engine, download the GLB and import it the way you would a Blender export.

Modular kit pieces that must snap on an exact grid, mechanical parts with precise dimensions, and anything that needs clean edge loops for deformation are better modeled than reconstructed. Summer's fabricate workflow lets an agent script your own Blender install for that kind of asset, and a generated mesh can go through the same step for decimation or a UV unwrap.