AI Image Generator
Type a prompt, pick a style and a canvas size, and get a real PNG back in well under a second. Rendered on this server — no API key, no queue.
Image generation runs through pluggable engines, exactly like the 3D studio. This deployment ships a local procedural engine that always works and produces real images. External image engines activate only when their API key is configured.
External image engines are not enabled on this deployment. No external image engine is configured, so generation runs on the built-in procedural engine. It produces real PNG output — patterns, gradients and surface textures — driven by your prompt.
Prompt
Describe what you want and the generator draws it: a repeating pattern, a colour gradient, a noise-built landscape, a tileable surface texture or a blueprint drawing. Five styles, four canvas sizes, one prompt box.
Rendering runs on the local procedural engine, which is always available. That is a design decision rather than a limitation: image generation is a pluggable engine slot, and external image models are adapters that switch on when the operator configures an API key. None is configured on this deployment, so nothing here pretends to call one.
How it works
Write the prompt
The prompt decides everything. Words like forest, ocean, fire, gold, metal or night steer the palette into the matching hue family, and the whole sentence is hashed into the seed that drives the composition. A prompt with no colour word at all still works — the hash picks a family on its own.
Choose a style and a canvas
Pattern for geometry, gradient for smooth colour, terrain for landscape, texture for a material surface, blueprint for a technical drawing. Match the canvas to where the image is going: 1280×512 for a banner, 1024×576 for a slide, 576×1024 for a phone screen, 768×768 for a square asset.
Generate, download, or send it to 3D
The PNG downloads straight from the result panel. If you want a model rather than a picture, send the same file to the studio: it becomes the reference for image-to-3D, and its colour is carried into the generated material.
AI Image Generator
Five genuinely different renderers
Pattern, gradient, terrain, texture and blueprint are separate renderers, not filters over one image. The terrain style builds a real value-noise heightfield and shades it; the texture style builds a surface with grain and pores; the blueprint style draws a grid and a set of line figures.
Colour that follows the prompt
Six hue families keyed off your words — greens for forest and leaf, blues for ocean and water, oranges and reds for fire and sunset, greys for metal and steel, yellows for gold and sand, deep blue-violet for night and space — plus a hash-derived family when no keyword matches. Every palette is five harmonised hex colours.
Reproducible bytes
The same prompt, style and size always produces the same file, byte for byte. The seed is a SHA-256 of the prompt rather than Python's process-salted hash, so a render you liked yesterday is the same render tomorrow.
Texture that tiles
The texture style wraps its noise lattice across both axes, so the edges meet exactly. Drop the result into a material as an albedo or roughness map and the repeat is invisible.
Four canvas sizes
768×768, 1024×576, 576×1024 and 1280×512 — square, landscape, portrait and wide, sized for social cards, slide decks, phone wallpapers and banners.
Numbers with every picture
Each render reports what it actually contains: the distinct colour count, the mean luminance, the seed and the palette family it landed in. Two renders can be told apart without opening them.
Image tool questions
Is this a real image generator?
Yes. Every pixel is computed from your prompt with numpy and encoded as a PNG by Pillow. There is no stock library and no placeholder: change one word in the prompt and the output changes.
Which AI model does it use?
None on this deployment. The built-in local engine renders your image on this server. Image engines are pluggable, so an external model can be added later with an API key — and the engine picker will name it when that happens.
Why do two similar prompts give different colours?
The hue family comes from keywords, but the exact hues, the composition and every other random choice come from a hash of the whole prompt. 'Forest' and 'a misty forest at dawn' land in the same green family and still produce different images.
Can I get exactly the same image again?
Yes — repeat the prompt, style and size exactly. Determinism is a hard guarantee of the local engine, and it survives restarts because the seed is derived with SHA-256 rather than a process-specific hash.
What does the texture style do that the pattern style does not?
Pattern is flat geometry with hard edges, which is what backgrounds and mockups want. Texture builds a surface: broad tonal variation, fine grain, shaded pores and mineral flecks, wrapped so it can be used as a material map.
Can I use the output as a reference for 3D?
Yes, and it is the shortest path on the site: generate the PNG here, send it to the studio, and image-to-3D turns it into a mesh. The local 3D engine reads the reference image's colour, contrast and aspect ratio, so the palette you chose shapes the material on the model.
Do Chinese prompts work?
Yes. The palette lexicon carries Chinese keywords alongside the English ones: 森林 and 树 pick the greens, 海洋 and 水 the blues, 火 and 日落 the oranges, 金属 the greys, 金 and 沙 the yellows, 夜 and 星 the deep blues. The seed is computed from the raw text, so a Chinese prompt is exactly as reproducible as an English one.
Is there a limit on how many images I can make?
No per-image limit and no queue. Rendering is local arithmetic: a 768×768 frame takes a fraction of a second and nothing is metered.