Getting started with Image Studio

The Image Studio layout

Three panels, left to right:

1. Inputs - your prompt, reference images, and any negative prompt
2. Settings - the model picker, aspect ratio, style controls, number of images
3. Results - generated images appear here. Click the Save button on any image (or "Save All") to store it to your gallery. Generated images are NOT saved automatically; if you leave the page without saving, they are lost.

Your first image

1. Open Image Studio.
2. Pick a model. If unsure, start with Nano Banana 2 Lite (fast, cheap, high quality for general work) or Flux 2 Pro (best photorealism).
3. Type a prompt. Describe the subject, style, lighting, and composition.
4. Set aspect ratio - 16:9 for wide, 1:1 for square (Instagram, album art), 9:16 for vertical (stories, mobile wallpaper).
5. Click Generate. Most models return in 5-15 seconds.

Prompting that actually works

Short punchy prompt structure:

[subject], [style], [lighting], [mood], [technical details]

Example:

"A woman with neon-pink hair sitting in a rainy Tokyo alley, cinematic photography, soft backlight from a vending machine, melancholy, shallow depth of field"

Good habits:

• Lead with the subject
• Add one style reference (cinematic, anime, illustration, oil painting)
• Specify lighting if it matters (golden hour, harsh overhead, neon glow)
• Only add technical terms (35mm, f/1.4) if you know what they do

Bad habits:

• Naming real people or artists
• Writing a paragraph-long narrative
• Stacking contradictory styles

Model cheatsheet

Live costs are in the model picker - they sync with the backend constants in server/image-routes.ts.

Not every model is from the same supplier. The Qwen models run on Alibaba and Z-Image and the Spicy models run on AtlasCloud, which means they keep working during an outage at our main provider. If a model you normally use is greyed out, one of these is usually still available.

Using reference images

Most models accept a reference image to influence style or composition:

• Style transfer - provide an image with the look you want, the AI generates your subject in that style
• Character consistency (Nano Banana 2) - provide multiple reference images of a character, generate new scenes with the same person
• Pose or composition - some models (Flux Kontext, when available) use the reference to lock in pose
• Editing an existing image - Qwen Image Edit Plus takes up to 3 references and is the cheapest edit path

Upload references directly or pick from your gallery. All references are stored on our R2 CDN - no external URLs will break.

Where your images go

Generated images show up in the Results panel but are NOT saved automatically. To keep an image, click Save on the image tile (or Save All to save the whole batch). Saved images land in your Gallery. Anything you don't save is lost when you leave the page or generate a new batch.

From your Gallery you can:

• Download (uses blob download, your session stays intact)
• Animate with an image-to-video model - see video-studio/getting-started
• Share to community feed
• Use as album art for a song

Common first-time mistakes

• Generating at low resolution for album art. Use 1:1 at the highest resolution the model offers. A 512x512 image looks rough on a streaming platform thumbnail.
• Not re-rolling. Image gen is stochastic. If the first result is meh, regenerate with the same prompt - variation is huge.
• Over-editing prompts. If you are 20 tweaks in and each result is worse, step back and try a cleaner shorter prompt.

When generations fail

See generation-failures. TL;DR: if credits come back, it worked as designed. If they did not, report the generation ID.