A beautiful, open-source desktop app for AI image & video generation.
Generate, iterate, organize — all in one place.
ImageStudio is a native desktop app for macOS and Windows that lets you generate images and videos using the best AI models — all through a single, polished interface. No browser tabs, no subscriptions, no clutter. Just you, your prompts, and your creations.
You bring your own fal.ai API key — one key for images and video — paying only for what you use. ImageStudio supports five image models, so you can compare results side by side.
Download the latest release for your platform from the Releases page.
| Platform | File | Notes |
|---|---|---|
| Windows | .exe installer or .zip portable |
SmartScreen may warn on first launch — click "More info" → "Run anyway" |
| macOS | .dmg disk image |
Unsigned — see below |
macOS note: ImageStudio is not signed with an Apple Developer certificate. macOS will block it on first launch. To fix this, open Terminal and run:
xattr -r -d com.apple.quarantine /Applications/ImageStudio.appThen open ImageStudio normally from your Applications folder.
git clone https://github.com/ibimspumo/ImageStudio.git
cd ImageStudio
npm install
npm run dev # Development with hot reload
npm run build:mac # Build distributable .dmg (macOS)
npm run build:win # Build distributable .exe (Windows)When you first open ImageStudio, a settings dialog appears. Paste your fal.ai API key — it covers image generation, video generation and reference uploads. You're ready to go.
fal.ai's image endpoints only accept URLs, so every reference image is uploaded to fal.ai storage before the request goes out. This is automatic — there is nothing to configure.
- Uploads are cached by content hash: the same image is transferred once, no matter how many models or how many images of a batch reference it
- Upload status is shown on the loading placeholders ("Uploading references…")
- The cache is cleared when you change your API key
- Works across all features: generation, chat, zoom out, collections
Settings has an Updates panel wired to this repository's GitHub Releases:
- Check now — asks GitHub whether a newer version exists (also runs quietly ~8s after launch, which you can turn off)
- Download — fetches the installer with a live progress bar and transfer rate. Downloads never start on their own
- Install — on Windows the app restarts into the new version; on macOS the disk image opens so you can drag the new build into Applications. macOS builds are unsigned, and macOS refuses to let an unsigned app replace itself, so that last step is manual
Type your prompt into the floating prompt bar at the bottom. It overlays the gallery so your images get maximum screen space. Press ⌘/Ctrl Enter to generate.
The prompt bar gives you full control over your generation:
- + button — attach reference images (moves above the text when images are attached)
- Model selector — choose which AI model to use (select multiple to compare)
- Aspect ratio — the ratios the selected model actually supports; a custom ratio is mapped to the closest one each model can produce
- Resolution — the output sizes the selected model offers (hidden for models with a fixed size)
- Image count — generate up to 4 images at once
- Quality — the quality tier of the OpenAI models (only shown for those)
- Background —
auto/transparent/opaque, shown only for GPT Image 1.5, the one model with the field.transparentreturns a PNG with a real alpha channel - Reference fidelity — GPT Image 1.5's
input_fidelity;treukeeps the reference's detail,freilets the model reinterpret it. Only active with reference images attached - Seed — lock a seed for reproducible results (hidden for the OpenAI models, which have no seed)
- Style preset — append predefined style suffixes to your prompt
- @ button — open your asset collections
- Queue — batch processing queue status
- Clear (⊗) — clear the current prompt, attachments, and collection references
- ⚙ — settings & about
Everything is non-blocking. You can fire off multiple generations and keep prompting while they render.
ImageStudio runs five image models, all through fal.ai. Click the model selector in the prompt bar to switch between them:
| Model | Provider | Aspect ratios | Resolution | Reference images | Seed | Price |
|---|---|---|---|---|---|---|
| GPT Image 2 (default) | OpenAI | 11 standard ratios | 1K–4K, quality-tiered | up to 16 | no | $0.005–$0.40 per image |
| GPT Image 1.5 | OpenAI | 1:1, 3:2, 2:3 only | three fixed sizes, quality-tiered | up to 16 | no | $0.009–$0.20 per image |
| Nano Banana 2 | 15, incl. 4:1 and 8:1 | 0.5K–4K | up to 14 | yes | $0.08 at 1K, ×1.5 at 2K, ×2 at 4K | |
| Nano Banana 2 Lite | 15, incl. 4:1 and 8:1 | fixed 1K | up to 14 | yes | ~$0.048 per image | |
| Nano Banana Pro | 11 standard ratios | 1K–4K | up to 14 | yes | $0.15, ×2 at 4K |
GPT Image 1.5 is the only model that can return transparency. Its background field takes
auto, transparent or opaque, and transparent gives you a PNG with a real alpha channel —
which is what Logo Mode is built on. It has no aspect ratio and no resolution axis at
all: the endpoint accepts exactly 1024 × 1024, 1536 × 1024 and 1024 × 1536, so any other ratio is
mapped onto the nearest of those three. It also exposes input_fidelity on its edit endpoint, which
controls how literally a reference image is preserved.
Every control in the prompt bar reflects what the selected model genuinely accepts — options a model does not have are hidden rather than silently ignored. None of the five support a negative prompt, so that control does not exist for images. Pick several models at once and the strictest limits apply, while values a given model cannot take are mapped to its nearest supported one.
Select multiple models using the checkboxes, then generate. ImageStudio fires off a request to each model in parallel, so you get results from all of them at the same time. Great for finding out which model handles your prompt best.
Click the aspect ratio button to choose from 10 presets — each shown as a visual box so you can immediately see the shape. Need something custom? Use the custom ratio input at the bottom with a live preview. Both fields are required before you can apply.
Presets: 1:1, 3:4, 4:3, 2:3, 3:2, 9:16, 16:9, 5:4, 4:5, 21:9
Custom: Enter any width:height ratio (e.g. 7:3) and click Apply.
ImageStudio supports AI video generation via fal.ai. Switch to video mode using the mode toggle in the top bar, then select a start frame image, write a motion prompt, and generate.
| Model | Provider | Durations | Notes |
|---|---|---|---|
| Seedance 1.5 Pro | ByteDance | 4–12s | Default. Good quality, affordable |
| Kling v3 Standard | Kuaishou | 5s, 10s | Fast, reliable |
| Kling v3 Pro | Kuaishou | 5s, 10s | Higher quality, supports negative prompts |
- Add a start frame — drag an image from the gallery onto the video prompt bar, click +, or use the file picker
- Describe the motion — write what should happen (camera movement, action, animation)
- Configure — choose model, duration, resolution, aspect ratio, and audio toggle
- Generate — press Generate or ⌘/Ctrl Enter
Videos are saved as MP4 files. The gallery shows both images and videos together, filterable via the toolbar. Videos play automatically on hover in the grid view.
- Full video playback with controls
- Details panel shows model, duration, estimated cost, and all metadata
- Save as MP4 directly (no format conversion needed)
- Image-specific features (chat, crop, inpaint, zoom out) are hidden for videos
Live cost estimates are shown in the prompt bar before generating. Costs are calculated from the model's per-second rate and your selected duration. After generation, the estimated cost is saved and displayed in the lightbox details.
fal.ai returns no billing data with a generation — neither the queue response nor the client exposes
billable units — so every figure ImageStudio shows is computed from fal's published list prices, and
is marked ≈ accordingly.
- Before generating — the prompt bar (and the chat input) shows what the pending request will cost. With several models selected, each one is priced separately and the total is the sum; hover for the per-model breakdown.
- After generating — the estimate is stored on the image and shown in the lightbox details.
- Running total — the title bar shows today's spend, with the all-time total on hover. Images generated before cost tracking existed carry no figure and are excluded (the tooltip says how many).
The rates live in src/shared/image-models.ts next to each model's capabilities, priced off what was
actually sent — a resolution a model cannot do is clamped before the request and billed at the
clamped tier.
The third mode, next to Image and Video, is built for one job: YouTube thumbnails. It removes decisions instead of adding them — what a thumbnail technically has to be is no longer a setting.
| Aspect ratio | 16:9, no selector |
| Generated at | 2K (Gemini) / 1920 × 1088 px (GPT Image 2) |
| Exported at | exactly 1920 × 1080 JPEG, under 2 MB |
| Models | GPT Image 2 (default), Nano Banana 2, Nano Banana Pro — Nano Banana 2 Lite drops out (fixed 1K) |
No model produces exactly 1920 × 1080: the Gemini endpoints return 2752 × 1536 (ratio 1.792) and fal.ai snaps GPT Image 2 to multiples of 16, so a requested 1080 comes back as 1072. Thumbnail mode therefore asks GPT Image 2 for 1920 × 1088 and normalises every export to a centre-cropped 1920 × 1080, stepping JPEG quality down until the file fits YouTube's 2 MB limit.
Every generation carries a thumbnail ruleset the user never has to retype: format and safe zones, the one-idea rule, faces and emotion, contrast and the 60/30/10 split, text limits, and click psychology. It ships as system_prompt on the Gemini models and is prepended to the prompt on GPT Image 2, which has no such field.
- Style —
Automatischis the default and adds no style instruction at all, leaving the prompt in charge. The three explicit steps areClean(editorial, typography-led, no arrows),Balanced(clear emotion, one accent colour) andMrBeast(expression at the limit, red arrow, outlined text, hyper-saturation) - Faces true — with reference images attached, an identity-preservation block is added: facial geometry, skin tone, hairline, glasses stay exactly as in the reference; expression and lighting may change, the identity may not. Faces are rendered large, unobstructed and never covered by text
- Video title — goes in as context with the explicit rule that any text in the image must not repeat it
Text in the image is not a separate control — write it into the prompt (Text: "30 TAGE") and the rules take care of the rest: 4 words maximum, spelled exactly as given, never across a face, never in the outer 5 %.
A project is a video: a title, optionally an angle, a colour. Thumbnails generated in the mode are filed under the active project automatically, and can be moved later by dragging them onto a project pill or via the card's move menu. Projects are a separate axis from workspaces — they only exist inside thumbnail mode, so eight videos in progress never leak into normal image work. Collections, @-mentions, drag & drop, the queue and export all behave exactly as elsewhere.
The preview (the ▶ button on a thumbnail card) renders the image in the surfaces it will actually appear in — search result, feed between two other thumbnails, home grid, "up next" sidebar, and mobile — in YouTube's light and dark themes, with an editable video title so title and thumbnail can be judged together. Four checks can be layered on top:
- Safe zones — the outer 5 % that some surfaces crop, and the corner the duration badge covers
- Thirds — rule-of-thirds guides
- Grayscale — does the image separate without colour?
- Squint — blurred: does it survive the fast scroll?
Below that the same image is shown at 120 × 68, 88 × 50 and 64 × 36 px. What is unreadable there does not exist in the feed. The export button writes exactly 1920 × 1080 and reports the resulting file size.
The fourth mode, next to Image, Video and Thumbnail, exists for one reason: GPT Image 1.5 is the only model in the app that can return a real alpha channel. A logo on a white square is a picture of a logo; a logo on transparency is a logo you can actually place.
| Model | GPT Image 1.5 — derived from the registry: a model needs a background field to appear here |
| Background | transparent by default, switchable to opaque or auto |
| Format | PNG, always. JPEG has no alpha channel, so it is not offered |
| Sizes | 1024 × 1024 (1:1), 1536 × 1024 (3:2), 1024 × 1536 (2:3) — the three the endpoint accepts |
Every generation carries a logo ruleset the user never has to retype: one memorable idea, flat
vector language, legibility filled in solid black at 16 px, centred with margin, two colours at
most, type only when the request names words, and no mockups, boards, frames or watermarks. With a
transparent background a second block is added that forbids painted backgrounds, glows, halos and
drop shadows — the things that quietly fill an alpha channel back in. GPT Image 1.5 has no
system_prompt field, so the block is prepended to the prompt.
| Type | What changes |
|---|---|
| Automatisch | No style block at all — the prompt alone decides (default) |
| Minimal | Geometric primitives on a grid, one colour, negative space doing real work |
| Wortmarke | The name itself is the logo; considered letterforms, no symbol beside it |
| Emblem | One closed outer form — badge, shield, seal — with everything locked inside it |
| Maskottchen | A single flat-vector character with a strong silhouette |
A transparent logo never goes through the JPEG anti-detection pipeline — that pipeline's first
and last step is JPEG, which has no alpha channel, and its middle step resamples the image twice,
which would blur exactly the hard edges a mark lives on. Transparent results are instead re-encoded
once as PNG through the browser's own encoder, which strips the generator's metadata and leaves
every pixel intact. Save, export, copy and drag & drop all hand out that PNG with its transparency
— Save in the gallery and the lightbox follows the stored file's extension, so it never writes a
.jpg.
Logos appear on a checkerboard in the gallery and the lightbox, in two mid greys rather than the usual white pair, so a white mark stays as visible as a black one. Chats opened from a logo inherit its transparency, so an iteration does not silently gain a background halfway through.
Logo mode filters the gallery to what it produced and keeps the normal workspace bar — a logo is an ordinary asset that happens to have an alpha channel. Collections, @-mentions, drag & drop, the queue and export behave exactly as elsewhere.
Click the + button in the prompt bar or drag & drop images directly onto it. When images are attached, they appear as thumbnails above the text field with a small + to add more. Remove all images and the + returns inline. These references are sent alongside your prompt for image-to-image editing — style transfer, face swaps, composition matching, etc.
Hover over any gallery image and click the crop icon, or use "Crop as Reference" in the lightbox. This opens a full-screen crop tool where you can draw a selection on the image. The cropped area is added as a reference to your prompt bar — perfect for isolating a face, texture, or detail from an existing generation.
Create named groups of reference images (e.g. @brand-photos, @product-shots). Type @ in the prompt to mention a collection inline. The images are automatically prepared and attached. Removing the @-mention chip from the text also removes the collection reference — nothing gets sent that you don't see.
Collections with more than 5 images are intelligently composited into grid layouts to stay within API limits.
Want to see what's beyond the edges of an image? In the lightbox, use the Zoom Out buttons (1.5x, 2x, 3x, 4x) to extend your image outward. ImageStudio creates a canvas with the original image centered and black borders, then sends both the canvas and the original as references — so the AI knows exactly what to fill. The result appears as a new image in your gallery.
Change the aspect ratio of any image directly from the lightbox. Select a target ratio from the grid — each button shows a visual preview with the source ratio nested inside the target ratio, so you can see exactly how the image will be extended.
- 9 target ratios — all standard ratios excluding the current one
- Model selector — defaults to the model that created the image, but switchable
- Same resolution — uses the original image's resolution setting
- The original image is sent as a reference with the new aspect ratio, and the AI extends the composition naturally
Upscale any image to a higher resolution directly from the lightbox. Available options depend on the original resolution:
- 1K images → Upscale to 2K or 4K
- 2K images → Upscale to 4K
- 4K images → Already at max resolution
Choose which AI model to use for upscaling via the dropdown (defaults to GPT Image 2). The original image is sent as a reference with instructions to recreate it at the target resolution while preserving every detail.
Note: Resolution output depends on the AI model. If the API returns a smaller image than requested, ImageStudio automatically detects the actual dimensions and corrects the resolution label in your gallery metadata.
Click the palette icon in the prompt bar to open the full-screen Canvas editor. Paint a color-coded sketch using multi-layer drawing tools, then let AI turn it into a detailed image.
- Brush — freehand drawing with adjustable size and color
- Eraser — remove strokes from the active layer
- Shapes — rectangle, circle, and line tools with optional fill
- Color Picker — 24-color palette + custom hex input
- Layers — up to 8 layers with visibility toggle, opacity slider, and drag-to-reorder
- Undo/Redo — up to 30 steps of history (⌘Z / ⌘⇧Z)
- Aspect Ratio — choose canvas dimensions before you start
Write a text prompt at the bottom and generate. Your sketch is automatically attached as a reference — the AI follows your composition, shapes, and color layout.
For precise control, switch to Expert mode. An intelligent panel on the right lets you:
- Detect Colors — automatically find all unique colors on your canvas
- Describe each color — a text field per color to explain what that region represents (e.g. "#FF0000 = red sports car")
- Attach references — add images or @-mention collections per color field for visual guidance
- General description — overall scene prompt that applies to the whole image
- Model/Resolution/Count — full control over generation parameters
All color descriptions are assembled into a structured prompt. Collections are deduplicated across fields — even if you mention the same collection in multiple color descriptions, it's only sent once.
Canvas-generated images support the Compare with Original feature. Your sketch is saved to disk automatically, so you can use the slider or side-by-side view to see your original sketch next to the AI result.
| Shortcut | Action |
|---|---|
B |
Brush tool |
E |
Eraser tool |
R |
Rectangle tool |
C |
Circle tool |
L |
Line tool |
⌘/Ctrl Z |
Undo |
⌘/Ctrl ⇧ Z |
Redo |
Escape |
Close canvas |
Select any image in the lightbox and click Inpaint to open the mask editor. Paint over the area you want to change, then describe what should appear there using the full prompt bar at the bottom.
- Brush tool — adjustable size (5–100px), with undo and clear
- Full prompt bar — same prompt bar as the main app, with attachments, @-mentions, seed, presets, and multi-model support
- Reference images — attach additional images as visual guidance (e.g. "Replace jacket with @image1")
- Green overlay — the masked area is sent as a green highlight on the original, so the AI can visually see exactly what to edit
- Lineage tracking — inpainted images link back to their source, visible in the detail panel and usable for comparison
Compare an image with its original version using the Compare with Original button in the lightbox. Available for any image that was derived from another (via upscale, zoom out, inpaint, canvas generation, or chat).
- Slider mode — both images overlaid with a draggable vertical divider
- Side-by-side mode — 50/50 split view with model and resolution labels
- Resolution-independent — images are displayed at the same visual size regardless of pixel dimensions
- Canvas sketches — compare your hand-drawn sketch with the AI-generated result
Click the star icon on any gallery card or in the lightbox to mark an image as a favorite. Favorites are persisted across sessions and can be filtered in the gallery toolbar.
Add custom tags to any image from the lightbox detail panel. Tags support autocomplete from all existing tags in your gallery. Use tags to organize images by project, theme, or any category you choose — then filter by tag in the gallery toolbar.
The gallery toolbar appears above your images with powerful filtering options:
- Search — full-text search across prompts and tags
- Model filter — show only images from specific AI models
- Aspect ratio filter — filter by aspect ratio
- Date range — Today, This Week, This Month
- Favorites only — show only starred images
- Tag filter — filter by one or more tags
- Sort — newest or oldest first
- Smart albums — auto-generated album chips for quick access (by model, date, favorites)
All filters can be combined and cleared with one click.
Style presets append predefined style suffixes to your prompt. Click the preset selector in the prompt bar to choose one.
Built-in presets: Cinematic, Anime, Photorealistic, Oil Painting, Minimalist, Watercolor, 3D Render
Create your own custom presets via Manage Presets — each preset has a name, optional emoji icon, and a suffix that gets appended to your prompt on generation.
Add generations to a queue for sequential processing. The generate button includes a dropdown to "Add to Queue" instead of generating immediately.
- Queue panel — slide-out panel showing all queued items with progress
- Sequential processing — one generation at a time, with progress tracking
- Persistent — queue survives app restarts
- Cancel & clear — cancel individual items or clear completed ones
When exporting images, metadata is embedded directly in the file:
- PNG — custom
tEXtchunks with prompt, model, seed, aspect ratio, resolution, and timestamp - Toggle "Embed metadata" in the export popover (on by default)
- Metadata is readable by standard tools like
exiftoolor file property viewers
The option disappears while Anti-Detection Processing is on — embedding the prompt would put back exactly what that step strips out.
Detectors for synthetic images key on the traces a generator leaves in the pixels: the periodic residue of its upsampling stack, an unnaturally clean frequency spectrum, PNG output that never went through a lossy encoder. Every generated image and thumbnail therefore runs through a short pipeline before it is written to disk — videos are never touched:
- JPEG at 95 % — quantises every 8×8 block, so no pixel keeps its exact generator value.
- 1 % squeeze per axis — the image is resampled down on X, then on Y, each pass on its own axis.
- Back to the original size and a closing JPEG round.
The result has the exact dimensions that were generated. Measured over 1000×1000, 1920×1088 and 2752×1536: roughly 86 % of all pixels change, at an average difference of 1.5 of 255 — invisible — and the whole pass costs 60–290 ms.
Because the pipeline runs before storage, everything downstream hands out the processed file:
gallery, export, copy to clipboard and drag & drop. Exported file names become neutral
(IMG_4831.jpg) instead of imagestudio-<uuid>.png; a thumbnail project keeps its own title,
which is your wording and gives nothing away.
Transparent images are the one exception. JPEG has no alpha channel and the squeeze step would
soften the hard edges a logo is made of, so an image generated with background: transparent is
re-encoded once as PNG instead — metadata dropped, every pixel kept, alpha intact. It stays a .png
through storage, export, clipboard and drag & drop.
Toggle it under ⚙ → Anti-detection processing (on by default). With it off, images are stored as PNG exactly as the model returned them.
Note: this changes the file, not the facts. Platforms that require you to disclose synthetic content — YouTube among them — expect that disclosure regardless of what a detector sees.
Want to iteratively refine an image? Hover over any image in the gallery and click the chat icon. This opens a conversation where each message builds on the previous result.
- The last generated image is automatically attached as a reference
- You can switch models between messages
- You can attach additional reference images — drop them onto the input or use the Images button
@-mentions work exactly as in the main prompt bar: type@to insert an attached image or a whole asset collection as a chip, and reference it in the prompt- Live cost estimate next to the keyboard hint
- All chat-generated images also appear in your gallery
When you open a chat from an image, that image's model, aspect ratio and resolution are pre-selected — an edit of a 9:16 image stays 9:16 unless you change it.
Click any image to open it in the lightbox. The info panel on the right shows everything about the image at a glance.
The info panel includes:
- Prompt — with a copy button
- Model — which AI model was used (shown as a readable name)
- Size — aspect ratio, resolution, and pixel dimensions (resolution auto-corrected to match actual image)
- Seed — with a copy button for reproducibility
- Duration — how long the generation took
- Date — when the image was created
- Tags — add/remove custom tags with autocomplete
- Chat origin — if the image came from a chat, click to reopen it
- Reference images — click to navigate to that image in the lightbox (if it's in your gallery)
- Favorite — star/unstar the image
- Reuse Prompt — restore the prompt, seed, @-collection mentions, and image references back into the prompt bar
- Start Chat / Continue Chat — open an editing conversation from this image
- Crop as Reference — select a region of the image to use as reference
- Inpaint — open the mask editor to selectively edit parts of the image
- Compare with Original — slider/side-by-side comparison with the source image (available for upscaled, zoomed, inpainted, canvas-generated, or chat-edited images)
- Aspect Ratio Change — regenerate the image in a different aspect ratio using AI
- Zoom Out — extend the image outward by 1.5x, 2x, 3x, or 4x using AI
- Copy — copy the image to your clipboard
- Save — quick export as PNG (with embedded metadata), or click the dropdown arrow to choose format and quality. Videos export directly as MP4
- Delete — remove from your gallery
Click the dropdown arrow next to "Save" to open export options:
- Format — PNG, JPEG, or WebP
- Quality slider — for JPEG and WebP, adjust from 10% to 100%
- Live file size — see the estimated file size update in real time
- Savings indicator — shows how much smaller the file is compared to the original
As your gallery grows, workspaces help you stay organized. Think of them as lightweight folders for your images.
- By default, you work without a workspace — all images are visible
- Click + Workspace below the title bar to create one
- When a workspace is active, new generations automatically go into it
- Move existing images between workspaces via the folder icon on hover
- Switch back to All to see everything
Each workspace gets its own color. Images in a workspace show a subtle colored bar at the bottom. Right-click a workspace tab to rename or delete it.
Press ? anywhere (outside a text input) to see the full shortcuts help overlay.
| Shortcut | Action |
|---|---|
⌘/Ctrl Enter |
Generate images |
@ |
Reference images/collections in prompt |
G |
Focus prompt editor |
⌘/Ctrl F |
Focus search bar |
F |
Toggle favorite (in lightbox) |
E |
Export (in lightbox) |
R |
Reuse prompt (in lightbox) |
⌘/Ctrl C |
Copy image (in lightbox) |
Delete |
Delete image (in lightbox) |
← → |
Navigate images in lightbox |
⌘/Ctrl Z |
Undo brush stroke (in inpaint/canvas) |
1–4 |
Open image 1–4 in gallery |
Escape |
Close topmost modal/dialog |
? |
Show shortcuts help |
| Layer | Technology |
|---|---|
| Framework | Electron + electron-vite |
| UI | React 19 + TypeScript |
| Styling | Tailwind CSS v4 |
| State | Zustand |
| Icons | Lucide React |
| AI (Images & Videos) | fal.ai API |
| Updates | electron-updater + GitHub Releases |
src/
├── main/ # Electron main process
│ ├── ipc/ # IPC handlers (generation, video generation, files, settings, metadata)
│ └── services/ # fal.ai image client, fal.ai video client, storage upload, image storage, updater
├── preload/ # Typed context bridge (window.api)
└── renderer/src/ # React UI
├── components/
│ ├── input/ # PromptBar, VideoPromptBar, ControlsRow, selectors, SeedInput, PresetSelector, CostEstimate
│ ├── gallery/ # Justified layout (row-based masonry), cards, GalleryToolbar, SmartAlbumBar
│ ├── canvas/ # Canvas editor: modal, workspace, toolbar, layers, color picker, expert mode
│ ├── chat/ # Image chat modal
│ ├── workspace/ # Workspace tabs and management
│ ├── collections/ # Asset collection manager
│ ├── presets/ # Style presets dialog
│ ├── queue/ # Batch queue panel
│ ├── tags/ # Tag input with autocomplete
│ └── shared/ # Lightbox, CropModal, InpaintModal, ImageCompare, ExportPopover, ShortcutsHelp, Settings, SpendIndicator
├── stores/ # Zustand (gallery, collections, chat, settings, workspace, crop, presets, queue, canvas, gallery-filter)
├── hooks/ # useImageGeneration, useVideoGeneration, useChatGeneration, useMentionEditor, useCanvasRenderer, useJustifiedLayout, useKeyboardShortcuts
├── types/ # API types, model definitions, shared interfaces
└── lib/ # Utils, image compression, anti-detection, date-utils, debounce, logger
Releases are built automatically via GitHub Actions when a version tag is pushed:
git tag v0.8.0
git push origin v0.8.0This triggers parallel builds on macOS and Windows, packages the app for both platforms (.dmg/.zip for macOS, .exe/.zip for Windows), and creates a GitHub Release with all artifacts and an auto-generated changelog.
- Your API keys are stored locally on your machine
- All generated images and videos are saved as files on disk — metadata is stored in lightweight JSON files (no base64 in memory)
- Generated images pass through the anti-detection pipeline before storage, which also drops any metadata the model's encoder wrote into the file
- ImageStudio never sends data anywhere except to fal.ai for generation, and to GitHub when checking for updates
- Reference images are uploaded to fal.ai storage because the image endpoints only accept URLs. Uploads are cached by content hash, so the same image is sent once.
- Existing data from older versions is automatically migrated on first launch
- No analytics, no tracking, no accounts
MIT — free and open source. Do whatever you want with it.
Contributions are welcome. Please open an issue first for major changes.