Controls and menus
Click any control’s bold name in the app for a one-line reminder.
Prompt and negative prompt
What you want, and what you don’t. Drag the small handle under a box to make it taller. See Writing prompts. Layout… opens the layout editor; while a layout is in use, the prompt box shows its description, and a click opens it.
Size
Width (W) and height (H) in pixels, always multiples of 16. The rotate button swaps them.
Aspect picks a shape (16:9, 3:2, 4:3, 1:1 and their portrait versions) and MP a size in megapixels. Both always show the current size: a shape that isn’t on the list shows as its ratio, such as 1.37:1, and then the MP menu offers it at ×1.5 to ×4 for upscaling.
MOSS-Image works best around 1 to 2 megapixels: proportions, hands, and lettering come out better there, even allowing for the extra pixels. Smaller sizes are fine for trying out ideas. Bigger pictures take longer: twice the pixels takes more than twice the time. See the size examples.
Images
How many pictures to make, one after another, with the same settings. Each gets its own seed.
Seed
The number the randomness starts from. 0 (the default) means a new random seed each time. The seed actually used is saved with every picture, so the same seed and settings give the same picture again.
Steps
How many refinement passes. More steps give cleaner detail and take longer. 20 is a good default and 30 is usually all you need; MOSS-Image’s authors recommend 50. See the steps examples.
Adherence
How literally the picture follows the prompt (often called CFG). 4 is the sweet spot, and anything from 3 to 6 gives much the same picture; much lower or higher values add artifacts. At 1, guidance is off: each step takes half the time, but the prompt is barely followed. See the adherence examples.
Output
The folder where pictures are saved, one sub-folder per picture. The folder button chooses another; Blokjutsu offers to create it if it doesn’t exist. The default is kept out of Time Machine backups and Spotlight; a folder you choose follows your Mac’s usual settings (Backups and Spotlight).
Start from image
Switch on to start from a picture instead of from scratch. See Working from a picture.
Variations
Switch on to make close variations of a picture. See Variations.
Advanced options
Leave these alone unless you’re curious. Reset advanced puts them back.
| control | what it does |
|---|---|
| Model | SFT is the polished model; Base is the raw one people train their own versions from |
| Sampler | the method that turns noise into a picture; see Samplers |
| SDE noise η | only for the Euler SDE sampler: how much fresh randomness each step adds |
| Shift | balances effort between overall layout (higher) and fine detail (lower); default 5 |
| Uncond skip, SLG layers, SLG scale, Skip interval | MOSS-Image’s built-in quality boosts; defaults are best |
| Max tokens | the longest prompt the model reads; longer prompts are cut |
| Size step | how much the W and H arrows change the size |
| Think first | Describe image and Enhance prompt plan before writing: slower, more accurate. Off each time Blokjutsu starts |
| Settings in PNG | also writes the prompt and settings inside each new PNG file; off by default, and asks you to confirm when turned on (more) |
| Low memory | sets the text encoder aside after each prompt, to save memory; on automatically below 48 GB (why) |
Samplers
The sampler is the method that turns noise into a picture, one step at a time. Steps is how many steps it takes.
| sampler | name in full | when to use it |
|---|---|---|
| Euler | Euler’s method, the simplest way of following a path in small steps | the default, and what MOSS-Image was tested with; best at 20 to 50 steps |
| Euler SDE | Euler’s method for a stochastic differential equation: some fresh randomness is added at each step | more varied texture, sometimes softer; SDE noise η sets how much randomness |
| DPM++ 2M | DPM-Solver++ (Diffusion Probabilistic Model solver), second order, multistep | far fewer steps: around 10 steps already gives a good picture, because each step also uses the one before |
| UniPC | Unified Predictor–Corrector: each step is predicted, then corrected | also for few steps; an alternative to DPM++ 2M |
DPM++ 2M and UniPC are general-purpose samplers rather than MOSS-Image’s own, so results can differ a little from Euler at the same seed. Examples shows all four side by side at 5 to 50 steps.
Describe image and Enhance prompt
Both use MOSS-Image’s own text encoder (the part that reads prompts) to write text. The first use asks to download a 2 GB extra piece of it.
- Describe image (on each picture, and in Start from image) writes a short, a medium, and a long description of a picture; see Describing images.
- Enhance prompt (above the prompt box) rewrites your prompt in three lengths.
Each version has Use as prompt and Copy under it. A description takes from a few seconds to a minute; with Think first on, several minutes.
Menus
Everything in the menus, menu by menu. Items marked macOS are added by the Mac to every app.
Blokjutsu
- About Blokjutsu: the version, credits, and license.
- Settings… (⌘-comma): preferences that aren’t image settings.
- Live preview: show the image forming while it generates.
- Image cache: memory used to keep history images ready for smooth scrolling.
- App data: the folder with Blokjutsu’s settings, logs, models, and (by default) your images, with a button to show it in Finder.
- Services (macOS): lets other apps act on what you’ve selected in Blokjutsu. Select some prompt text, for example, and Services offers what your Mac’s apps can do with text, such as making a note or searching the web. The list depends on the apps you have. Services Settings… opens System Settings, where the checkboxes turn each service on or off in every app’s Services menu and right-click menu. Blokjutsu doesn’t use Services itself.
- Hide Blokjutsu (⌘H), Hide Others (⌥⌘H), and Show All (macOS): hide or show windows.
- Quit Blokjutsu (⌘Q).
File
- Generate (⌘-Return): the same as the Generate button.
- Import Parameters… (⌘O): sets the controls from a
parameters.jsonfile, or from a PNG made with Settings in PNG turned on. Use it to repeat or adjust an image someone sent you. - Show Output Folder in Finder: opens the folder your images are saved in.
- Show App Data in Finder: opens the folder with Blokjutsu’s settings, logs, and models.
- Download Model…: downloads the MOSS-Image model, or checks for and fetches only the files that have changed. It asks before anything is downloaded.
- Download Captioning Head…: downloads the extra piece that Describe image and Enhance prompt need. It asks before anything is downloaded.
- Delete All History…: permanently deletes your images and everything made from your prompts. See Your history.
Edit
These work in whichever text box you’re typing in.
- Undo (⌘Z) and Redo (⇧⌘Z): step back and forward through your typing. After Use as prompt, ⌘Z brings back your own prompt.
- Cut (⌘X), Copy (⌘C), and Paste (⌘V). Copy also works on text you’ve selected on an image’s card, such as its prompt.
- Select All (⌘A).
- Start Dictation… (macOS): type by speaking.
- Emoji & Symbols (macOS): insert special characters.
View
- Appearance: System follows your Mac’s light or dark setting; Light or Dark keeps Blokjutsu that way whatever the Mac uses.
- Enter Full Screen (macOS, ⌃⌘F).
Window
- Minimize (⌘M): sends the window to the Dock.
- Zoom: makes the window fill the screen without going full screen; choose it again to restore it.
- Anything else in this menu, such as arranging windows on the screen, is added by macOS.
Help
- Search (macOS): finds any menu item by name.
- Blokjutsu User’s Guide: this guide, in your web browser.
- MOSS-Image User Guide, MOSS-Image Technical Report (arXiv), and MOSS-Image on GitHub: the model authors’ own pages, in your web browser.
- Show Log: opens Blokjutsu’s log file. If something goes wrong, the full error is there; include it when reporting a problem.