Generating and editing images with ISH: what to expect
ISH includes image generation and image editing tools for users who want to turn prompts or uploaded images into visual outputs. The experience is meant to be simple: choose the image tool, write a prompt, upload a reference if needed, and generate.
But image models are not magic vending machines. Quality, speed, limits, and supported features depend on the active provider and model. This guide explains how to use the feature without overselling it.
What you can do
Depending on the model available in ISH at the time, image tools can help with:
- generating a new image from a text prompt
- editing parts of an existing image
- restyling a reference image
- creating visual variations
- producing square, portrait, or landscape outputs
- turning rough ideas into draft artwork
The best use is creative iteration. Generate a draft, inspect what worked, then adjust the prompt. A second prompt is usually better than trying to make the first one perfect.
Choosing a prompt
A strong prompt usually includes:
- the subject
- the setting
- the style or medium
- the composition
- lighting or mood
- what to avoid
For example:
A clean product-style photo of a black ceramic coffee mug on a walnut desk, soft morning light, shallow depth of field, no logo, no text, no watermark.
That is better than:
make a cool mug image
Specific prompts reduce random output and make editing easier.
Editing existing images
When editing, describe the exact change and what should stay the same.
Good:
Replace only the background with a quiet studio backdrop. Keep the person, face, clothing, pose, and lighting unchanged.
Weak:
make this better
Image editing works best when the instruction is narrow. If you ask for a broad style change and object edits at the same time, the model may change more than you wanted.
Limits and reliability
Image generation can fail, slow down, or return a different style than expected. That can happen because of provider capacity, safety filters, unsupported inputs, or model-specific limitations.
Do not assume every model supports the same image sizes, editing behavior, or number of reference images. If one request fails, simplify the prompt first. If it still fails, try a different model or a smaller edit.
Copyright and style safety
Do not ask models to copy living artists, branded characters, copyrighted scenes, or private images without permission. For blog banners, product art, and social previews, the safer approach is to ask for original visuals with no logos, no readable text, no watermark, and no copyrighted characters.
That is the same rule we use for BLOGish artwork. Original images are cleaner for publishing and much safer for long-term use.
Practical workflow
- Start with a text-only prompt.
- Generate one draft.
- Identify the biggest issue: subject, lighting, style, framing, or details.
- Rewrite the prompt around that issue.
- Use editing only when you need to preserve part of an existing image.
- Save the version that actually fits the page or project.
Good image generation is mostly good direction. The clearer your instruction, the less time you spend fighting random output.
Sources
- OpenAI image generation docs: https://platform.openai.com/docs/guides/image-generation
- Google AI image generation docs: https://ai.google.dev/gemini-api/docs/image-generation
- Pollinations: https://pollinations.ai/
