Prompting
Describe what you want in natural language. The agent interprets, refines, and generates. Good prompts tend to include:- Subject — what’s in the image
- Style — photography, illustration, 3D render, editorial
- Mood — minimal, dramatic, warm, cinematic
- Technical — lighting direction, camera angle, color palette, lens
Choosing the model
Image and Video Generator nodes show the model name before you generate. A new node starts on Memo’s current default model, which the picker lists once under its own name, first in Featured; it is not an automatic choice based on your prompt. Open the model picker to choose a specific model. Its available settings and credit estimate update before you run it. The settings panel shows the same model name. Existing saved projects keep their settings.Using references
Drag any inspiration from your library onto the Studio canvas. When you mention it in chat, the agent uses it as visual context — matching its style, composition, or color palette. If a connected agent has written a visual analysis of the inspiration through the connector, the agent gets the full creative breakdown — not just the pixels. It can match the lighting setup, use the extracted palette, or apply the composition style to a new subject. Your team’s comments on it ride along too: Send to Studio lists them in the chat’s first message as direction.Brand assets as constraints
Type@ to mention a brand kit or specific asset. The agent loads your brand context and generates within those constraints. This is how you get consistency without repeating yourself every prompt.
See Using Brand Assets in Studio for the full workflow.
A fixed Image Generator sends the prompt you write and the images you attach. Adding an image from a Brand Kit does not automatically add its rules, character descriptions or reference instructions. Inline mentions become their visible names; image names and provenance remain available with the saved result. Creative Brief and chat runs continue to use their selected guidance.
Aspect ratio and output size
Choose the format and the model’s real output size. 1K / 2K / 4K are provider resolution tiers, not ratings of image quality. Available sizes depend on the selected model. Where a model offers both 1K and 2K, Studio’s standard is 2K: it is preselected, used when you don’t choose, and priced that way. Pick 1K for cheaper drafts or 4K for print. A fixed-size model keeps its own size. GPT Image has separate Output size and Rendering effort controls. On GPT Image 2.6 and 2.5, the sizes are real pixel dimensions in any shape up to 3:1: 1K is about 1 megapixel (1024 × 1024 square), 2K about 4 megapixels (2048 × 2048 square, 2720 × 1536 at 16:9) and 4K the provider’s maximum (2880 × 2880 square, 3840 × 2160 at 16:9). The size picker shows the exact dimensions for your shape. Effort runs from low to max: it changes detail and price, not the number of pixels. Older saved settings keep Provider preset (about 1 megapixel) until you choose a size. GPT Image 2.6 Flare is Memo’s default model; a new image starts at 2K with medium effort. Sunburst is its precision sibling: the same price, more fine detail, and about three to four times slower. On fal, a 1K High Flare image took about 16 seconds and a 2K High one about 38 seconds; Sunburst took about 57 seconds at 1K High. Krea 3 has a Thinking control next to the size. Medium works with or without references and is priced by output size. High and Ultra are for edits: they need at least one reference image and stay greyed out until you add one. High plans the edit more carefully, keeps untouched areas and corrects text, for a fixed amount more per image. Ultra is Krea’s deepest mode: it always renders at 2K, keeps the shape of your reference (the shape picker shows Auto), and is priced by the number of references rather than the size. On Krea 3 each reference image also adds to the price; the credit cost on the Generate button always includes it. Muse has no size toggle. It chooses dimensions from the aspect ratio at roughly 2.5 megapixels; a square output can be below 2048 pixels. Inspect the delivered file dimensions when a minimum size matters. Provider tiers and explicit size requests do not guarantee correct lettering, logos or hardware geometry.Working with a result
Click a finished image or video on the canvas and a small toolbar floats above it with the actions you reach for most:- Edit prompt — opens the result in the editor, with the original shown beside the prompt for comparison: rewrite the prompt, manage references in the References card (the dashed + tile opens the picker), change the model, aspect ratio, quality or duration in Output, then press Regenerate (the credit cost is on the button). The new version lands as its own node next to the original, which stays untouched; the editor says so under the button
- Reframe (images) — re-generate the same picture at another aspect ratio and quality
- Status — mark the result Needs Review, In Progress, Approved, Rejected or Pending; the status shows next to the node’s name and is saved with the project
- Duplicate and Delete. A duplicate sits to the right of the original with a clear gap and is named with the next number (“Hero shot #2”)
Cinema presets
Cinema presets turn one image into a full nine-shot coverage plan — camera coverage, closeups, cutaways, or story beats, organized by category. Right-click any image on the canvas, pick Cinema Presets, browse or search the tree, click a preset, then choose how many shots to run (3, 6, or 9 — each shot bills as its own image) and a quality (1K / 2K / 4K). Each preset generates separate high-fidelity images — one per shot, grouped together on the canvas — rather than a single collage, so no shot loses detail. Every shot is its own full-resolution image you can reframe, save, or turn into video individually. Behind the scenes:- Memo analyzes your image’s subject and rewrites each shot description around it
- Each shot generates independently on a high-fidelity model, using your image as the visual reference
- Running fewer than nine samples evenly across the plan, so 3 shots still span it from wide to detail
- Shots land as a named group arranged three per row, matching the preset’s plan
Extract Grid
Right-click any reference image on the canvas and pick Extract Grid to reuse its layout as a template. Memo analyses the source — composition, panel structure, alignment — and turns it into a reusable grid you can populate with new content. Useful when you’ve found a layout you like and want to keep its bones while changing everything inside it.Iteration
After a generation, you can:- Ask for variations — “try a warmer color palette”
- Request edits — “remove the text overlay”
- Start fresh — “let’s try a completely different direction”
Generations run in the background
Generations are queued. The moment you submit, the request is accepted and the canvas shows a placeholder node. If the chat started it and the node lands out of view, the canvas moves to it. While it waits, the node shows where it is: Queued · #2 while it waits its turn, then Generating with the time passing. You can keep prompting, switch projects, or close the tab — when you come back, completed generations are already there. There’s nothing you have to wait around for. If a generation does not come back, the chat says so: its card turns into an error with the reason, and Memo adds one line saying what happened and what to press. A picture that fails is not charged. A picture still out after five minutes (a video after an hour) may yet land on the canvas; if it never does, nothing is charged and the held credits return on their own. Press Retry on its card on the canvas to run it again.Uploading your own references
Three ways to bring pictures in, as many at a time as you like: Dropping a picture straight onto the chat panel also works: it uploads, appears as a thumbnail chip on your message, and the agent can use it as a reference. The picture is kept only with that project until you decide what it is: the small bookmark on the chip offers Save to Inspirations (a reference you want to find again) or Add to Brand Kit (brand material such as a product shot or a logo). Up to six pictures a message.- Drag images from your desktop onto the canvas. They land where you drop them.
- Press Cmd+U (or the upload button in the toolbar) to pick files.
- Paste a copied image with Cmd+V.
Design Editor
Click into any image node and select Edit to open the Design Editor — a lightweight overlay tool for adding text, simple shapes, and brand-locked color fills directly on top of an image. The AI Design Assistant on the right can do two kinds of change:- Text, shapes and color fills are drawn on top of the image. They are free.
- Changing the picture itself (a warmer background, a new style, removing an object) runs an image generation with your image model. Each edit costs that model’s image price, with your current design attached as the reference. The editor shows the price before you send, and again while the edit runs. The result appears in the project’s generations like any other image, and an edit only runs when the workspace has the credits for it.
- Instagram Post — 1080×1080
- Instagram Story — 1080×1920
- X / Twitter Post — 1200×675
- LinkedIn Post — 1200×627
- TikTok Cover — 1080×1920
- Original — keeps the source resolution
Video generation
The agent can also generate videos with control over:- Duration — from 2s up to 30s in a single pass, depending on the model
- Resolution — 480p up to 4K on models that support it
- Aspect ratio — 16:9, 9:16, 1:1, and more on select models
- References — a start frame, an end frame, and reference images, where the model supports them; the composer shows a slot for each input the selected model accepts
- Creative direction — describe motion, camera movement, and action
Naming and finding generations
Your image is usable as soon as it finishes. Memo then saves a short description and a search index so you can find it by meaning, such as “navy mug ad.” This background indexing does not change the image or add a credit charge. After saving the description, Memo starts indexing immediately; a scheduled worker recovers interrupted or failed jobs. A temporary indexing delay does not prevent opening, downloading or reusing the image by its direct reference. Every generation gets a title automatically, but the ones you’ll reuse deserve real names. Open any generation and use the Name card in its detail panel to rename it — names are what the search bar matches, so a keeper namedproduct-fw25-sneaker-red-v2 is findable next week, next month, by anyone on the team.
The same names work for agents: a connected agent can name outputs at generation time and re-find them by name later, so a library your team names by hand and a library your agent builds stay one library. Two habits keep it clean: name outputs the moment they’re approved (drafts can keep their automatic titles), and never reuse a name — a new version gets a new suffix.