Images
Generate, edit, upscale and cut out images inside a conversation. Tyron owns the image tools; Ashley, Lucy and Steve put the results to work.
Ask for a picture, get a picture — inside the conversation, not in another app.
Unyo generates images from a description, edits them when you say what's wrong, upscales them, and cuts out the background. The results drop straight into a social post, a document, a website or your library. And because your products and your brand kit live in the Neural Core, the image usually already looks like your business rather than a stock photo.
Who does what
The image tools belong to Tyron, your audiovisual agent. This isn't a style preference — it's a runtime access control. The other nine agents have image.generate, image.modify, image.upscale and image.remove_background stripped out of their toolbox before they ever see them.
| Agent | What it can do with images |
|---|---|
| Tyron | The only agent with the four image tools: generate, edit, upscale, remove background |
| Ashley | Cannot call the image tools. Her post draft carries a generated image, and she edits it through the post card's own image commands |
| Lucy and the 8 other document agents | Illustrate a document — generate, upload, pick from your library, or find a real web photo |
| Steve | Imports an image you attached into the website project. No generation |
| Blake / Tyron | Charts produced by real Python in the sandbox come back as image cards |
Why Ashley can't paint
Ashley writes the visual brief; Tyron paints. Ashley is denied every image.* tool at the runtime layer, so no agent can quietly do another's job. You still get an image on your post — see Images in social posts below for how.
Generate an image
- Open a conversation with Tyron.
- Describe what you want. Be specific about the subject, the setting and the mood — vague prompts give vague images.
- Add an aspect ratio if the default square doesn't suit you.
- The image appears as a card and the panel opens on it automatically.
- Not right? Say what's wrong. Don't start over — see Edit by talking.
Generate a photo of my hazelnut praline jar on a dark walnut table, morning window light from the left, shallow depth of field, a few whole hazelnuts scattered around. Landscape.
Aspect ratios
image.generate takes one of six ratios. The default is square.
| Value | Typical use |
|---|---|
square | Instagram feed, avatars, thumbnails |
portrait | Vertical feed, stories-shaped visuals |
landscape | Blog headers, website heroes |
feed_portrait | Tall social feed posts |
feed_landscape | Wide social feed posts |
ultrawide | Banners, wide section backgrounds |
Aspect ratio is a steer, not a pixel spec
Ask for landscape and you get a landscape image — but Unyo does not guarantee an exact pixel size like 1200×630. The ratio is passed to the image model as a constraint, not enforced afterwards by a cropper. If you need exact dimensions for an ad slot or a template, crop the result yourself after downloading it.
Edit by talking
You don't need to re-prompt from scratch. Tell Tyron what to change and he edits the image you're both looking at — if you don't name an image, he takes the last one in the conversation.
Things that work well as edit instructions:
- Light and colour — "warmer", "less contrast", "cooler shadows"
- Composition — "pull back a bit", "move the product to the right third"
- Add or remove — "take out the fork", "add a linen napkin under it"
- Restyle — "same shot, but as a flat illustration with my brand colours"
Upscale to HD
Ask Tyron to upscale and the image goes through a conservative upscale pass — more resolution, no reinvention of the subject. Use it on an image you've already approved, as the last step before print or a large hero.
Remove the background
Ask for a cut-out and you get the subject on transparency — ready for a website hero, a slide, or compositing onto your brand colour.
Product-aware generation
This is the part that makes the images look like yours.
If you name a product that's saved in your Neural Core and that product has photos, Unyo automatically feeds those real photos to the generator as references, locks onto the product, and applies the visual style you saved for it. You don't have to ask for any of this — matching the product name in your prompt is enough.
- Save the product in your Neural Core with at least one real photo.
- Name it, literally, in your prompt.
- The generated image keeps your actual product rather than inventing a lookalike.
Name it exactly
The match is on the product name appearing in your prompt, and the longest matching name wins. "Praliné Noisette 250g" matches; "the praline thing" doesn't. Products synced from Shopify are protected from accidental edits, so they stay a reliable reference.
Style references and your brand kit
- Your brand kit — ask for your logo or your colours and Unyo pulls the real assets: the light logo first, the dark one as fallback. If neither exists, the agent asks you instead of guessing.
- A saved style image — reference a style or mood image saved in your Neural Core by name and it's used either as the image itself or as the style to reproduce on your subject.
- An attached photo — attach an image in chat and it's analysed for description, style, mood and key elements, so what's written about it describes what's actually in it.
Use the "Studio Minimal" reference from my Neural Core — same lighting and palette — but with the 250g jar as the subject.
Images in social posts
Ashley's post draft arrives complete: caption, hashtags, platform toggles, and the image already generated. The image is produced on the server during the run, not lazily when you open the card — which is why it renders identically live and after a page reload.
From the card you can:
- Edit an image in place — "make it brighter" edits it and keeps its position in the carousel
- Add an image — appended as a style-coherent companion, using the previous image as a visual reference
- Replace an image — "I hate it, redo it" deletes and regenerates
- Target a specific image or, in a stack of drafts, a specific post
One image unless you ask for a carousel
A normal post request produces exactly one image. Carousels are only built when you explicitly ask for one (2–10 images) — the model is instructed never to invent one. Carousel variants are forced to differ: each uses a different shot directive (telephoto, macro, wide environmental, flat-lay, in-use with a person, seasonal staging), and a similarity check rejects and retries a variant that came back too close to the reference.
The image that comes with a post draft is generated square. If you need another ratio for that visual, ask Tyron for it directly and bring the result into the post.
Images in documents
Nine of the ten agents write real documents, and any of them can illustrate one four ways:
| Mode | What it does |
|---|---|
| generate | Creates an illustration from a prompt |
| library | Picks from your saved images |
| search | Finds a real web photo — and re-hosts it, so the illustration never turns into a broken hotlink |
| upload | Uses an image you provide |
Add an illustration to the "How it's made" section — the roasting drum, warm light, documentary style.
Images on a website
Attach an image in a Steve conversation and ask him to use it: he imports it into the project at a path you choose (assets/hero.jpg, for example). Steve has no generation tools — he places images, he doesn't paint them. Images up to 3 MB.
Your library and inline galleries
- Browse — ask to see your recent images and you get an inline gallery of thumbnails you can search by the prompt that made them.
- Save a post's images to the Neural Core — say the word and the post's images are copied into a colour-coded folder in your knowledge base, each one keeping its generation prompt as its description. They then become reference material for future generations.
- Web photo galleries — ask a visual question and Unyo can show a gallery of real web photos inline. The gallery only appears when at least three genuinely relevant images are found; otherwise you get a written answer instead of padding.
What it costs
Priced in credits, per action:
| Action | Credits | ≈ |
|---|---|---|
| Generate an image | 150 | €0.15 |
| Edit an image | 150 | €0.15 |
| Upscale to HD | 30 | |
| Remove the background | 20 | |
| Browse or search your library | 0 | free |
| Look at your brand kit | 0 | free |
| Import an image into a Steve site | 5 |
A failed image doesn't cost you
Every one of these tools is refundable on error. If the generation fails, the credits go back automatically — and an hourly janitor refunds anything left stranded. On top of unit costs, each run bills the real cost of the model's tokens, so a short request costs less than a long one.
Troubleshooting
The image ignored my product. The product name has to appear literally in your prompt, and the product needs at least one real photo saved in the Neural Core. Check both, then ask again using the exact saved name.
The dimensions aren't what I asked for. Expected — the ratio is a constraint on generation, not a post-crop. Crop the download to your exact spec.
I asked Ashley for an image and she talked about Tyron. Correct behaviour. She doesn't hold the image tools. Either work from the post card's own image commands, or open a Tyron conversation for standalone visuals.
No gallery appeared when I asked a visual question. Web photo galleries need at least three relevant results. Fewer than that and Unyo answers in words rather than showing you weak matches.
My logo didn't get used. Check that your brand kit actually has a logo saved. Unyo tries the light logo first and falls back to the dark one — if neither is there, the agent will ask you for it rather than invent one.
Which agent do I ask for an image?
Tyron, for anything standalone. He is the only agent holding image.generate, image.modify, image.upscale and image.remove_background — the other nine have them removed from their toolbox at the runtime layer, not just discouraged in a prompt. That said, you rarely need to think about it: Ashley's post drafts arrive with an image already, and the document agents can illustrate what they write.
Can I get an image at an exact pixel size?
Not guaranteed. You choose from six aspect ratios — square, portrait, landscape, feed_portrait, feed_landscape and ultrawide — and that ratio is passed to the generator as a constraint. It is not enforced by a cropping step afterwards, so exact dimensions like 1080×1350 are not promised. Download and crop if you need to hit a precise slot.
Will the image actually show my real product?
Yes, if the product is saved in your Neural Core with photos and you name it in your prompt. Unyo detects the name, passes the real photos in as references and applies any visual style you saved for that product — automatically, without you asking. If the product isn't saved or isn't named, you get a generic interpretation instead.
How many images come with a social post?
One, unless you ask for a carousel. A carousel is 2–10 images and is only built when you explicitly request it. When you do, each image is pushed to a genuinely different shot — different framing, distance and staging — and near-duplicates are rejected and regenerated.
Do my images stay available after the conversation?
Yes. Generated images are saved and browsable from your library, searchable by the prompt that created them. You can also copy a post's images into a folder in your Neural Core, where each keeps its prompt as a description and becomes reference material for future generations.
What happens if a generation fails?
You're refunded. All four image tools are marked refundable on error, so a failure returns the credits automatically, and an hourly cleanup refunds anything that got stranded. You're never charged for an image you didn't receive.