Free tool
Models like Seedream respond best to structured prompts: subject first, then style, lighting and camera. Build one with chips, copy it, call the API.
a lone lighthouse on a cliff above the sea, cinematic film still, golden hour light, wide establishing shot
Try it in the playground →Structured prompts (subject → style → lighting → camera) follow the pattern that diffusion models are trained on, and usually beat single-sentence prompts.
Same structure the chips compose — subject, style, lighting, camera, palette — written out longhand. Paste these straight into a generation tool or an API call.
Product shot
“Product photo of a matte black wireless earbud case on a travertine pedestal, softbox studio lighting with gentle top-left key, 85mm macro, shallow depth of field, warm neutral palette, commercial advertising style”
Poster with text
“Minimalist music festival poster that says "NIGHT SIGNAL" in bold condensed white letters, deep indigo background with a single neon waveform line, flat vector style, high contrast, generous negative space at the bottom for a date line”
Cinematic landscape
“A lone lighthouse on a basalt cliff above a storm-grey sea, matte painting style, golden hour backlight breaking through clouds, wide establishing shot at 24mm, muted teal and amber palette”
Need a model to run it against? Browse the catalog or grab an API key.
Structure and specificity. Lead with one concrete subject, then layer style, lighting, camera and mood in that order — models weight early tokens more heavily. 'A lone lighthouse on a cliff, matte painting style, golden hour backlight, wide establishing shot, muted teal palette' will beat 'beautiful lighthouse picture' on every frontier model.
The templates here follow the conventions of modern diffusion models — Seedream, Nano Banana and GPT Image 2 all respond to them. GPT Image 2 handles long, literal instruction-style prompts best. Try the same prompt across models to see the style differences.
Not on modern instruction-following models. Instead of 'no blur, no distortion', state the positive: 'sharp focus, clean anatomy'. Syntax like (--no) or (word:1.2) belongs to older Stable Diffusion workflows and is ignored or mishandled by the models on our gateway.
One to three sentences for photoreal work, up to a short paragraph for complex scenes with text or many elements. Past that, returns diminish and later clauses start contradicting earlier ones. If you can't say what changes between two prompts in five words, they're too long.
Yes — prompts are yours, and images you generate through our API or tools are yours to use commercially, subject to our terms and each provider's usage policies.