Making images with CSuite.
Every control, menu, setting and tool in the image workspace, with a figure for each, followed by recipes that chain them into finished work. The short version lives on the main guide; this page is for when you want to know exactly what a button does.
Get around
The workspace
Open Image from the rail. Like the other workspaces it is a file list on the left, a large preview in the middle and the prompt box at the bottom; what is specific here is the editing toolbar across the top of the preview and the AI menu beside it.
- 1 · Sidebar
- Your project folder pill, then every image in the folder as a thumbnail row, newest first. While a generation runs a “Generating…” row sits at the top until the file lands.
- 2 · AI menu
- Upscale 2×, Upscale 4× and Remove background, run on the image you have open.
- 3 · Editing toolbar
- Resize, Crop, Adjust, Rotate, Annotate, Quality and Format. Each opens an options row under the header and turns the preview into a work canvas.
- 4 · Info
- Dimensions, size, format, modified date and, for images CSuite made, the model and prompt behind them.
- 5 · Model button
- The selected image model by name. Click for the picker and that model’s settings; hover for a one-line summary including the active aspect ratio and format.
- 6 · Preview
- The selected image fitted to the panel, with a zoom pill in the corner. With an image selected on a model that takes references, it is also the source of the next generation.
- 7 · Composer
- The prompt, the wand, Add reference, the active settings and cost estimate, and Create (which reads Update when a selected image is the source, or Upscale on the upscaler model).
The ✕ beside the folder pill (“Clear selection”) deselects the image so the next prompt creates a new one instead of revising it. Esc does the same when your focus is not in a text field, and so does clicking Image in the rail again.
Files and thumbnails
- Thumbnails are made on your machine for every file, SVG included. A tile pulses while it is being made and shows a broken-image glyph if the file cannot be read.
- More actions on a row: Show in Finder (Show in File Explorer on Windows, Show in file manager on Linux), Duplicate, Copy image (onto the clipboard as a PNG, ready to paste into a deck or a message) and Delete, which asks first. Delete or Backspace with a row selected also deletes.
- Add your own images by dragging PNG, JPG, WEBP, GIF or SVG files from your desktop onto the list (“Drop images to add them here”). They are copied into the project folder and behave exactly like generated images: editable, usable as references, and open to the AI tools.
- Filter and sort appear above 20 files: a “Filter images by name…” box and a sort button cycling Newest, Oldest and Name.
- Names are given by CSuite: a generated image takes the first two words of its prompt, title-cased (“Espresso Grinder.png”), a revision of a project image is “source-v2-prompt”, and a clash gets a “ 2”, “ 3” suffix rather than overwriting anything. Rename in the title field above the preview.
Generate
Models and settings
The Model button at the top right opens the picker. Cloud models run through Runware, Replicate and OpenRouter on your platform keys, or direct from a vendor such as OpenAI, Google, xAI or Black Forest Labs on that vendor’s key. Local models run through the bundled GGML runtime on any supported machine, and through Ollama on macOS. The pick is kept for the image workspace alone. Browse every model with its samples, prices and limits in the catalog.
Auto
Leave the picker on Auto and CSuite reads the prompt and picks: a logo or icon goes to a vector model, a quick draft to a fast one, a photoreal brief to a flagship. With several platform keys it prefers Runware, then Replicate, then OpenRouter; with none, an installed local model. Under Auto the popover shows the picker alone, the cost estimate is hidden, and a setting can ride in the prompt instead: “red car, 16:9” sets the ratio for that run, and the fragment never reaches the model. Auto never picks the upscaler, a vendor-only model, or a model you added yourself from a provider’s library; pick those by name.
Settings
- Aspect ratio, grouped Square, Landscape and Portrait; the choices are the ones the selected model accepts, so the list changes with the model. A model that edits an attached image may offer “match input” and follow the first reference’s shape.
- Format: PNG, JPEG or WebP where the model offers them; Recraft’s vector tiers return SVG.
- Size, Image size, Resolution or Quality: a model’s own output tiers (1K, 2K, 4K; standard or high), shown only when it has them. Larger tiers cost more.
- More settings holds the rest, counted in its label: a seed for repeatable results, step counts on local models, safety tolerance, background transparency and whatever else the model takes.
- Per-run inputs such as a negative prompt or a mask image appear beside the composer on the models that take them, since they belong to one run rather than to the model.
Values are remembered per model, so your Flux ratio and your Nano Banana ratio can differ. Hovering the Model button lists the active ones.
Generating
A failed run keeps its error above the box with a Retry button and your prompt in place. Every generated image remembers its model, prompt and settings; the Info popover shows them and can refill the composer.
References and revisions
Most cloud models, and Qwen Image 2.1 locally, take reference images: pictures the model edits or draws guidance from. When the selected model supports them an Add reference row appears in the composer with the model’s limit, and its hover text says what an attached image does for that model.
- Revise what you have: select an image in the sidebar on a model that takes references. It becomes the source automatically, counts as one of the reference slots, and the button changes to Update. Describe only what should change; the result is saved as a new version beside the original, never over it.
- Guide a new image: clear the selection and attach references instead. A product photo plus “the same bottle on a beach at sunset” keeps the bottle.
- Limits are the model’s: the add button disables at that model’s cap and the row reads “Reference limit reached”. Nano Banana 2 on Runware also takes a Ref. video. Ollama’s image models take no references, so pick the same model on GGML to edit.
- Attachments clear when you change model, because each model’s slots differ.
Uploaded references are staged, not added to your project folder, and never name the output. Chat and Workflows can run the same edit on an image you hand them.
The wand
The wand at the start of the composer’s toolbar rewrites a short prompt into a full one: subject, style, lighting, composition, mood. It runs on your text model, not the image model, and is off until you have typed something. The rewrite goes back into the box for you to read and send.
It knows what it is writing for. With a reference or a selected source attached it writes an editing instruction: what to change, where, and what must stay, without describing parts the model cannot see. On a vector model it drops photographic vocabulary that SVG cannot carry. On a model with its own negative prompt field it keeps exclusions out of the main prompt, and a model with a character cap gets a rewrite that fits.
Upscale and remove background
The AI menu at the start of the preview header runs three jobs that need no prompt on the image you have open: Upscale 2×, Upscale 4× and Remove background. The result is saved as a new file beside the original and becomes the selection.
- They run on Runware or Replicate: the one you have selected, or the first connected one when you are on Auto, OpenRouter, a vendor key or a local model. They are billed like a generation, to your platform account.
- With neither connected, the menu still opens and an action shows a message with an Open Models button that takes you to the providers screen.
- One action at a time per image: a second request is refused with a message. While one runs the menu leads with Stop.
The upscaler model
For enlargements beyond 4×, or with detail and realism controls, pick P-Image Upscale in the picker. It takes no prompt: the box reads “No prompt needed. Attach the image to upscale.”, the reference row reads “Attach the image to upscale”, and the button becomes Upscale once an image is attached (a selected sidebar image counts). The Upscale factor setting enlarges each side by that factor and the popover explains it inline. It runs on Runware or Replicate and is priced by the output size, so a 16× run on a large source costs more than a 2× on a thumbnail; a factor that would pass the platform’s pixel cap is refused with that message.
Local generation
Z Image Turbo, Flux 2 Klein, Qwen Image and Qwen Image 2.1 run on your own machine through the bundled GGML runtime (built on stable-diffusion.cpp) on Apple silicon, Windows x64 and Linux x64; on macOS, Z Image Turbo and Flux 2 Klein also run through a second bundled Ollama. Download a model under AI model providers (weights can be paused, resumed or cancelled, and shared parts download once), and it appears in the picker without a restart.
- First run shows “Installing runtime…”, “Downloading model…” and “Loading model…” before “Generating…”. After that the weights are on disk and everything runs offline, with no key and no per-image cost.
- Steps is the one dial, offered as a short ladder per model (Z Image Turbo 4, 8 or 12; Flux 2 Klein 2, 4 or 8; Qwen Image 10, 20 or 40). More steps, more detail, more time. The aspect ratio redistributes the model’s native pixel budget rather than adding pixels.
- Editing locally: Qwen Image 2.1 is the one local model that takes references (up to three) and can revise a selected image; it is licensed for non-commercial use. Ollama’s image models take none.
- Memory: on a machine where the weights are large for the GPU, the runtime offloads to the CPU and tiles the decode automatically; it is slower, not broken. The sidebar run row and Stop work as for a cloud run.
View and edit
Zoom, info and copy
- Zoom: the pill in the preview’s corner steps from 10% to 1600%, with 1:1 for actual pixels and Fit (the default, which never enlarges a small image).
⌘/Ctrl+ scroll or a pinch zooms around the pointer; once zoomed, scroll or drag to pan. A new file or a saved edit returns to Fit. - Transparency shows over a checkerboard for formats that can carry it (PNG, WebP, GIF, AVIF, SVG). A file that cannot be decoded says “This image could not be loaded.” in place of the preview.
- Info (the i at the end of the header) lists Dimensions, Size, Format and Modified, plus Model and Prompt for images CSuite made. Use prompt & settings refills the composer: the prompt always, the settings when that same model is still selected.
- Copy image on the row menu puts the image on the clipboard as a PNG (“Copied to clipboard. The image is ready to paste.”), whatever format the file is.
Every tool below shares three things: an options row under the header, a Save a new copy switch (on, the result is a sibling file; off, the file is overwritten in place), and Apply / Reset. Zoom is locked while a tool is open (“Finish or cancel the edit to zoom”), because the tool measures against the fitted image. All of it runs on your machine; nothing is uploaded. SVG and GIF files open for viewing only; use Format to make a PNG first.
Resize
The Size menu has three kinds of entry:
- Custom: type W and H in pixels, with Lock aspect to keep the proportion as you edit one. Values run from 1 to 16384.
- Heights from 240p to 1080p: the width follows the source’s aspect ratio (“aspect ratio preserved”).
- Publishing presets: OG link · 1200×630, IG square · 1080×1080, Story · 1080×1920 and X header · 1500×500. These are fixed frames, so the image is scaled to cover and centre-cropped (“center-cropped to fit”), never stretched. Crop first if the subject is off centre.
Crop
Pick a Ratio (Free, or any of the app’s ratios such as 1:1, 16:9, 9:16, 4:3, 3:2 and 21:9) and drag the marquee over the preview. Drag inside it to move it, drag a corner to resize, and the thirds grid helps you place the subject. Apply crop writes the result; a marquee smaller than the minimum is refused (“Crop area is too small”). The same ratio list is used by the video workspace’s crop tool.
Adjust
Three sliders, Brightness, Contrast and Saturation, each from 0 to 200% with 100% as the unchanged middle. The preview updates as you drag, and Apply bakes exactly what you see into the pixels. Zero saturation makes a greyscale version; Reset returns all three to 100%.
Rotate and flip
Left and Right turn the image 90°; Flip H and Flip V mirror it. Stack as many as you like before Apply: a turn plus a flip saves as one edit. Rotation changes the pixel dimensions (a 1280×716 still becomes 716×1280), which the Info popover reflects after the save.
Annotate
Five modes: Brush (click and drag to draw), Arrow (drag from tail to tip), Rectangle (drag to outline a region), Text (type in the field, then click the image to place it) and Eraser (drag to erase marks). Color and Size apply to the next mark; Size also sets the text height. Marks sit on a layer over the untouched photo until Save marks, so Undo removes the last mark and the eraser never touches the pixels underneath. Saving flattens the marks into the file, or into a new copy.
Format and quality
- Format re-encodes the image as PNG, JPEG or WebP: same pixels, a different container. PNG keeps transparency; JPEG flattens it onto white; WebP is the smallest for the web. This is also how an SVG or GIF becomes an editable raster.
- Quality appears for JPEG files only and sets the compression from 1 to 100%. Around 80% is small enough for mail without visible loss; the Info popover shows the new file size after the save.
Recipes
A product shot, start to finish
A logo as real SVG
Clean up a photo
Mark up a render for feedback
Free drafts, paid finals
Reference
Keyboard shortcuts
⌘ is Ctrl on Windows and Linux. The main guide lists every workspace’s shortcuts together.
Preview
⌘ + scroll- Zoom around the pointer (or pinch)
Scroll / drag- Pan a zoomed image
Esc- Clear the selection (outside a text field)
Files and composer
Enter / Shift+Enter- Create / add a line to the prompt
Delete / Backspace- Delete the selected image (asks first)
⌘+ / ⌘- / ⌘0- Zoom the interface in, out, or back to 100%
Troubleshooting
- “Upscale and background removal run on Replicate or Runware. Connect one under Models to use them.” The AI menu needs one of those two platforms whatever model you generate with. Press Open Models and add a key.
- “An AI action is already running on this image.” Wait for it, or press Stop at the top of the AI menu.
- No Add reference row. The selected model takes no references (Ollama’s image models, most local rows). Pick the same model on GGML or a cloud model.
- “Reference limit reached”. That model’s cap; remove one chip, or remember the selected sidebar image counts as a slot. Clear the selection to free it.
- Create reads Update and edits instead of creating. An image is selected and the model takes references. Press the ✕ beside the folder pill, or Esc, to clear it.
- No cost estimate. Shown only for a concrete cloud model with published rates; Auto and local models show none.
- “This image could not be loaded.” The file is damaged or in a format the preview cannot decode. The thumbnail glyph says the same. Re-export it elsewhere, or delete it from the row menu.
- “Finish or cancel the edit to zoom”. A tool is open; press Apply or Reset and close it.
- The editing toolbar is greyed for a file. SVG and GIF are viewable only. Run Format to PNG, then edit the copy.
- “Enter width and height between 1 and 16384.” Custom resize values are out of range.
- “Crop area is too small”. Drag a larger marquee before Apply crop.
- “Only image files (PNG, JPG, WEBP, GIF, SVG) can be added here.” Other files are refused by the sidebar; convert them first.
- A local model is missing on this machine. The GGML runtime has builds for Apple silicon, Windows x64 (CUDA) and Linux x64 (Vulkan); Intel Macs and arm64 Linux or Windows have none, and Ollama’s image runtime is macOS only. Use a cloud model there.
- A local run is very slow or the first result is flat grey. The runtime fell back to CPU offloading or ran out of GPU memory. Lower the steps, pick the smaller model, or close other GPU-heavy apps; a decode that ran out of memory is reported as a failure rather than saved.
- The upscaler refuses a factor. The output would pass the platform’s pixel cap; choose a smaller factor or start from a smaller source.
- Auto never picks the model I want. Auto skips the upscaler, vendor-only models and models you added from a provider’s library. Select it by name.
One image, every size
The resize presets are the publishing frames you actually use. The trick is the order: crop for the subject once, then let the presets cover and centre.