Give the AI pictures and shaders

Three ways to give the app more than words:

• Picture → 3D model. Start Generate 3D model from a photo, a picture in the Explorer, or a screenshot of the viewport.
• Pictures in the assistant's chat. Attach a picture and ask about it ("what is this?", "build something like this", "make this a 3D model").
• Your own GLSL. Write the vertex and fragment shader of a material yourself, in the shader editor's </> view.

Location data (EXIF, GPS) is removed from every picture on your device, before it is sent.

Try it

Templates ▸ Shader garden opens three objects already wearing your own GLSL: sliding stripes, a breathing glow and a rim light. Select one, open the shader editor and press </>.

Picture → 3D model

1. Set up a mesh provider: Settings ▸ Tools & connections ▸ Mesh providers. Meshy (hosted) takes pictures as it is. A ComfyUI workflow takes them when it has a {{IMAGE}} placeholder in a LoadImage node (see 3D Generation).
2. Add menu ▸ Generate 3D model… and pick the picture: From device…, Camera (on a phone), Viewport, or From Explorer…. You can also paste or drop a picture onto the window (a dockable window since 1.65, see The Generate window). In the Explorer, a picture's menu has Make a 3D model….
3. A prompt is optional. With Meshy's Refine mode it becomes the texture prompt.
4. Press Generate.

Before a picture goes to Meshy

Meshy is a third-party service. Before anything is sent, the app asks:

• what it costs: about 20 credits in Preview mode (untextured) or 30 in Refine mode (textured), from Meshy's price list. Meshy charges what it charges; the number is a guide.
• where the picture goes: to Meshy's servers, with a link to Meshy's privacy policy.

Cancel sends nothing. The same question comes up when the assistant makes a model from a picture you attached. A ComfyUI on this computer or your local network is your own server, so it does not ask; one anywhere else (a rented GPU, a hosted preset) asks before the picture leaves, and gets the picture under a plain name, never your file's.

Limits

Picture types	JPEG, PNG, WebP, HEIC (HEIC is converted on your device)
Size sent	up to 2048 px on the long side and 4 MB, re-encoded on your device
Location data	removed before sending

When it does not work

You see	What to do
No mesh provider set up yet	Press Set one up in Tools & connections.
… can't start from a picture	Use Meshy, or add {{IMAGE}} to a LoadImage node in your ComfyUI workflow.
Meshy rejected the API key	Check the key in Settings ▸ Tools & connections ▸ Mesh providers.
Meshy: out of credits	Add credits on meshy.ai, or switch the provider to Preview, which uses fewer.
… did not finish within 10 minutes	Try again, or try Preview mode, which is quicker.
… could not be reached	Check the connection and the provider's address.

Pictures in the assistant's chat

Press + in the assistant's composer (or paste, or drop a picture on the window). Up to 4 pictures a message, each sent as a JPEG of up to 1568 px, location data removed. The + menu says so before you pick anything.

Pictures only go to models that can see them:

Provider	Pictures
Gemini	yes
Grok	vision models (for example grok-4); grok-3 models cannot, and the + says so and names one that can
Custom (vLLM / OpenAI-compatible)	vision models by name (…-vl…, llava, pixtral, …) yes; an unknown name may try, and if the server refuses, the chat says so and carries on without the picture
OpenAI's or Anthropic's own endpoint as a Custom provider	not yet: their presets come in a later release

The assistant keeps the pictures of the newest two picture turns; older ones become a line saying a picture was there, so long chats stay small. Pictures in the assistant's chat stay on your device's conversation: other people in the session do not see them. Viewers cannot attach pictures.

"Make this a 3D model": with a picture attached, the assistant can start a picture → 3D job with it (the Meshy question above comes first).

Your own GLSL

Select an object, open the shader editor, and press </>. Three tabs:

• GLSL: the shader the graph compiles to, read-only, with Copy.
• Graph JSON: the graph as text. Edit it and press Apply; Reload throws your edits away.
• Your GLSL: a vertex + fragment pair that replaces the graph while it is on. Apply turns it on, Use the graph goes back (your text stays).

What you get:

• three.js built-ins: position, normal, uv, modelViewMatrix, projectionMatrix, normalMatrix, cameraPosition.
• uniform float uShaderTime; is the session's shared clock, so every peer sees the same frame.
• Any other float, vec2, vec3 or vec4 uniform you declare gets a field under the code. A Set Shader Uniform node can drive it from a behaviour.
• A compile error shows above the code with its stage and line ("Fragment line 2: …"). Press it to go to that stage. The last good shader keeps running while you fix it.

The rules (so a shader cannot freeze anyone's screen)

Loops	while and do are refused; a for loop declares its int counter and counts up to a number or a #define above it (up to 1024 per loop)
Work	one pixel (or vertex) may do about 32768 operations: each call, operator and assignment counts one, a loop counts its body once per step, a function counts what it runs, and both sides of a branch count. A matrix costs its size (a mat4 product counts 64, inverse 256), and so do whole arrays and structs. A 128-step raymarch is about 2000
Arrays	at most 64 elements, one dimension; reading one with a computed index counts its length
Parameters	no out or inout parameters: return the value instead
Size	64 KB per stage, plain ASCII outside comments, no backslashes or carriage returns
Preprocessor	#define NAME 32 (a name for a whole number, once) and two #includes, after your loops: <tonemapping_fragment> and <colorspace_fragment>. No #if, #ifdef, #undef or other directives
Textures	not in v1 (no samplers)
Lighting	none built in: a pair is unlit unless you light it yourself; skinned and instanced meshes do not follow it
Where	surface materials; post effects show the GLSL and Graph JSON tabs only

Before a shader runs, each device tests it on its own: it compiles and draws a small frame in a separate worker, with the shader's own field values and a running clock, then the first real frame is timed. A shader that fails, is too slow, or makes the graphics context drop is turned off on that device with the reason and a Try again button. The scene keeps the shader: it is that device's verdict, not everyone's, and the device remembers it until you press Try again.

Where your GLSL runs

Where	What happens
Your own scenes	runs, after the device's test
A game you export	runs for players, after their device's test. A shader that came from someone else is switched off in the export, and the export says so (open it in Your GLSL and press Apply to make it yours)
Someone else's shader, in a room you join, a community template or a scene file	runs only once your device's test has passed. If your browser cannot run the test, it stays off with a notice and Run it anyway
Quest	off by default, because one stuck shader can close the whole browser. Turn it on with Settings ▸ VR ▸ Display ▸ Turn on custom shaders (or the button in the </> view). Each shader is still tested first

Keyboard and screen readers

The </> tabs, the stage switch, Apply and Use the graph are buttons that take Enter; Tab moves from the stage switch into the code. Each code box has its own name ("Your GLSL, fragment shader", "Graph JSON", "Generated GLSL, read-only") and errors are announced.