An Local AI Studio Built for Imagining
SeasonStudio is an all-in-one AI creation studio that runs entirely on your own machine: write, translate, generate images, video, music and speech, and read what a picture holds.
Get the app:
Windows Store
Mac App Store
Later
Eight tools in one window, with nothing to sign into and nothing sent away.
New here? The eight tools ·
what each one is really good at yet ·
your first ten minutes.
Eight tools, one window, and none of them phone home.
Most creative AI lives behind a login and a credit counter. SeasonStudio puts the whole workshop on your desk instead: the models sit in a folder you chose, inference happens on your own CPU or GPU, and the results land as ordinary files. Work on a train, work on a plane, work on a client's confidential brief — the studio does not care, and never knows.
On your machine, by design
No account, no server, no analytics. Every generation runs locally through llama.cpp, stable-diffusion.cpp, GGML and ONNX Runtime. Turn the network off and everything still works.
One window, not eight apps
Text, translation, image, video, music, speech-to-text, text-to-speech and vision share a single shell, a single results panel and a single set of shortcuts. You stop switching tools mid-thought.
Your models, your shelf
One small language model ships inside so the app is useful the moment it opens. Everything beyond it is yours to import — point the studio at a folder of GGUF or ONNX files and it picks them up.
One afternoon, three worlds
A realist matte painting, a storybook illustration and an inked comic panel — the same machine, the same window, three different models on the shelf.
Image
Cinematic realism, generated locally at 1024×1024.
Image
Storybook palette, soft light, no reference sheet needed.
Image
Line art and flats, straight out of the sampler.
Prompts and model choices for these three are listed on the Image tool page.
The eight tools
Every panel follows the same rhythm: pick a model, describe what you want, watch it arrive, save it. Full descriptions →
Write and rewrite
Draft, summarise, restructure, or talk an idea into shape with a local language model.
Cross languages
Dedicated translation models rather than a general chatbot pressed into service.
Paint from a sentence
Text to image, image to image and editing, with the sampler previewing as it goes.
Move a still
Short clips from a prompt or from a single frame. The most demanding tool here.
Score and sound
Stable Audio models for loops, beds and effects when you need something under a cut.
Transcribe
Turn a recording, an interview or a voice memo into text you can actually edit.
Give it a voice
Read a script aloud, describe a voice in words, or keep a voice consistent across takes.
Read a picture
Caption, OCR, detection and segmentation for when the input is the image.
A prompt beside its result
There is no node graph and no scripting layer. You type a sentence, choose how long to let the sampler think, and the picture appears in the results panel.
- Prompt
- a cartoon wizard reading under a glowing mushroom, watercolour storybook illustration, soft rim light, fireflies, warm cream paper
- Avoid
- harsh contrast, text, watermark
- Model
SDXL_base- Size
1024 × 1024- Steps
28, guidance6.5- Backend
CUDA· roughly 20 seconds on a desktop GPU
Placeholder settings, shown to make the shape of a run concrete. Real prompts, models and timings vary with the shelf you build and the machine you build it on — see hardware expectations.
Where it runs
Two local inference stacks sit under the eight panels. Language and audio models run through GGML and llama.cpp in GGUF form; image, video and vision models run through ONNX Runtime. Each stack takes the fastest backend your machine actually has, and falls back to plain CPU when it has to.
| Platform | Language & audio (GGML) | Image, video & vision (ONNX Runtime) |
|---|---|---|
| Windows 10 / 11, x64 | CPU · Vulkan · CUDA | CPU · DirectML · CUDA |
| macOS, Apple silicon | CPU · Metal | CPU · CoreML |
| iOS · Android | Under consideration — see the roadmap. | |
A studio that runs locally is bounded by the machine it runs on, and pretending otherwise would only waste your afternoon. The hardware expectations table says plainly which tools are comfortable on a thin laptop and which really want a desktop graphics card.
From a sentence to a file
Every tool takes the same path, so once you have learned one panel you have learned all eight. Two safety gates sit in that path: one reads your prompt before any model is loaded, and one looks at the pixels before anything is shown on screen or written to disk.
Only what this run needs
The studio never asks you to provision a whole model zoo up front. A run names exactly the file it is missing, and nothing else is fetched. Keeping a lean shelf is a perfectly good way to use the app.
Results you can hand over
Output is ordinary PNG, MP4, WAV and text. There is no proprietary project format standing between you and the next tool in your pipeline, whatever that tool happens to be.
What is ready, and what is still learning
On-device generation is not uniformly mature, and the gap between modalities is wide enough to plan around. This is our own reading of where each tool stands today.
| Tool | Today | What that means in practice |
|---|---|---|
| Translate | Ready | Dependable across major language pairs. The most predictable tool in the app. |
| Image | Ready | Good enough for real illustration, concept work and stylised art. |
| Speech to text | Ready | Clean recordings transcribe well; heavy accents and crosstalk still need a pass. |
| Text to speech | Ready | Natural for narration. Prosody drifts on very long scripts, so cut them up. |
| Vision | Ready | Captioning, OCR and detection are solid on ordinary photographs. |
| Text | Ready | Small local models draft and edit well. They are not a research assistant. |
| Music | Early | Loops, beds and effects are convincing. Full arrangements are not there yet. |
| Video | Experimental | Short clips at modest resolution, slowly. Budget several attempts per shot. |
The reasoning behind each row, model by model, is on the Models page.
Nothing leaves the room
The most private studio is the one with nowhere to send your work. — the design brief, in one line
No account, no telemetry
There is no sign-in, no server component and no analytics SDK. Your prompts, imports and results stay on the file system you already own.
Content safety, on device
Two local gates sit around every run — one on the prompt, one on the pixels. Both run on your machine, like everything else here.
Downloads you asked for
The only network traffic is a model file you chose to fetch. Import from disk instead and the studio never opens a socket at all.
Licensing you can check
Every model the app can load is listed with its licence, so commercial use is a decision you get to make with the facts in front of you. See the list →
Get it
Windows and macOS
SeasonStudio is arriving on the Microsoft Store and the Mac App Store. Those store pages are not live yet — this section will carry the real buttons the day they open.
Free to start, one purchase to open it up
The app installs with a small language model already inside, so you can try the whole shell before deciding anything. A single one-time Premium unlock opens all models and hardware acceleration backends — there is no subscription. Current pricing is shown in the app, behind the Premium button.
Not sure it will run well on your machine? Start with hardware expectations.