Home › Guide
SeasonStudio is meant to be learned by using it, not by reading about it. This page exists for the handful of things worth knowing before you start, and the handful more worth knowing when something surprises you.
Nothing here takes longer to read than the model you are waiting on takes to load.
The app opens with a small language model already inside, so there is something to do immediately. A suggested order:
Resist the temptation to download the biggest model first. A fast model teaches you the app; a slow one teaches you to wait.
One window, three regions, and they never move. The tool list runs down the left, the controls for the current tool sit in the middle, and results collect on the right.
Eight entries — Text, Translate, Image, Video, Music, STT, TTS and Vision. Switching tools does not throw away what you had; you can go back to a panel and find your settings where you left them.
Bottom left. One screen, one price, a Restore action for a purchase you already made. It unlocks all models and the hardware acceleration backends; there is no subscription and nothing recurring.
Model, prompt, and the two or three numbers that matter for this tool. Panels deliberately expose less than a node-graph app would — the aim is a decision you can make in seconds.
Every output from every tool, newest first, with a full-size view a click away. Save what you want; the rest costs you nothing but a little disk cache.
This is the one piece of housekeeping the studio asks of you, and it works in the least demanding way we could find: a run tells you what it is missing, and nothing else.
Two image models beat five: one fast one for exploring, one heavy one for finishing. The same logic applies everywhere — a small transcription model for voice memos and a large one for interviews will serve you better than one compromise in the middle.
Image models run from a couple of gigabytes to twenty; video models are larger still. Check hardware expectations before starting a download you will not be able to run.
There is no secret syntax and no magic word list. Prompting well is mostly about being specific in the right places and quiet in the others.
Name the subject, then the light, then the medium. Everything else is decoration. — the shortest useful prompting advice we know
The Text panel is a good prompt editor. Describe the picture you want in ordinary sentences, ask for it as a compact image prompt, and paste the result into Image. This is also the direction the app itself is heading — see the roadmap.
Output is ordinary files in ordinary formats — PNG for images, MP4 for video, WAV for audio, plain text for text. There is no project format and no library database that has to be exported from.
Image generation draws a live preview while it samples. If the composition is wrong at step five it will still be wrong at step thirty — stop it, change one clause, run again. This single habit will save you more time than any hardware upgrade.
When something works, note the model, seed and steps somewhere before you move on. Persisting that automatically, so a run can be reopened and re-run later, is the first thing on the roadmap — until then, a text file beside the image does the job.
Two gates sit around every run, and both of them run on your machine like everything else in the app.
Both gates use a small classifier bundled with the app. Like the rest of the studio, they work with the network switched off, and they report nothing anywhere.
There is no account, no server component, and no analytics or crash-reporting SDK. Your prompts, imported files and generated results never leave your machine.
Downloading a model file you explicitly asked for, asking the store which purchases you own, and — only when you press Report — opening the report page in your browser. That is the entire list. Import your models from disk instead and the app has no reason to open a connection of its own at all — you can verify this by running it offline, which is a perfectly normal way to use it.
Models sit where you put them. Results and cached previews sit in the app's own folder on your disk. Settings are a local file. Deleting the app takes all of it with you.
The Premium unlock is handled entirely by the Microsoft Store, the Mac App Store or Google Play, which means we never see a payment detail. Restoring a purchase on a new device goes through the same store mechanism: the app asks the store what it owns.
Results carry a Report button. It opens this site's report page with the
record placed after the # in the address, a part browsers never send to a
server; that page has no analytics and no third-party scripts. Nothing reaches us until
you send the email or open the GitHub issue yourself, and a GitHub issue is public.
| What you see | What it usually means |
|---|---|
| Generation is far slower than you expected | The run has fallen back to CPU. Check the backend selection, confirm the Premium unlock if you are relying on an accelerated backend, and try a smaller model or a smaller output size. |
| A run stops or the app runs out of memory | The model does not fit. Drop the resolution, use a more heavily quantised model, close other applications, or move down one tier in hardware expectations. |
| A model will not load | Usually an incomplete download or a missing companion file — image models need a matching VAE, video models need their text encoder. Re-fetch the file and check the pair. |
| Your prompt was refused | The prompt gate declined it before any work started. Rephrase and run again; nothing was downloaded and nothing was recorded. |
| A result came back as a grey placeholder | The output gate replaced a frame that scored over the safety threshold. Adjusting the prompt is usually enough. |
| Images are blurry or washed out | Steps too low, guidance too high, or a size the model was not trained for. Return to the model's native resolution first, then adjust one number at a time. |
| A transcription is full of mistakes | Audio quality dominates here. Try a larger Whisper-family model, and if the recording has overlapping speakers expect to edit. |