Home › Tools
Eight tools, one window
Every panel in SeasonStudio is laid out the same way: choose a model on the left, describe what you want in the middle, watch it arrive on the right. Learn one and you have learned all eight.
Each section below answers the same four questions — what it is for, what it needs on your shelf, what it does well, and where it will let you down.
Text Ready
A local language model for the writing around the work: a first draft, a tighter second pass, a summary of a long brief, a list of names for a character you have not designed yet. It is the panel you will open without noticing.
On your shelf
A compact instruct model ships inside the app, so this panel works the moment you install. Larger models — an 8‑billion‑parameter class model, for instance — can be imported when you want more range.
Where it shines
- Rewriting your own rough sentences into something you would send.
- Turning a page of notes into a structured brief.
- Generating variations: ten titles, ten taglines, ten character names.
- Expanding a half-formed idea into a prompt for the Image panel.
Where it will not
A small model on a laptop is not a research assistant. It does not know today's news, it will invent a citation if you ask for one, and it reasons less reliably than the large hosted models you may be used to. Treat it as a fast, private writing partner rather than an oracle.
Translate Ready
Purpose-built translation models rather than a general chatbot pressed into service. Paste a paragraph, pick the direction, get a translation that holds its register instead of drifting into an explanation of itself.
On your shelf
Dedicated translation models in two sizes: a light one that keeps up with typing, and a larger one for text where nuance matters. Both are single files on disk.
Where it shines
- Major language pairs, in both directions, at paragraph length.
- Confidential material that must not be pasted into a web form.
- Localising your own store copy, captions and UI strings.
Where it will not
Poetry, wordplay and dense idiom still want a human. Rare language pairs are weaker than the common ones. If the source text is a mess, the translation will be a faithful mess.
This is, in our own testing, the most consistently dependable tool in the app — a good place to start if you want to see what on-device inference feels like.
Image Ready
The centre of gravity of the studio. Text to image, image to image, and instruction-led editing, with the sampler drawing a live preview so you can abandon a bad seed early instead of waiting out the whole run.
On your shelf
The whole open image ecosystem is available: the SD 1.5 line for speed, the SDXL and SD 3.5 lines for fidelity, plus newer families and dedicated editing models. Mix a fast model for exploring with a heavy one for finishing.
Three from one afternoon
Cinematic realism, dramatic rim light, 1024×1024.
Storybook watercolour, warm palette, soft edges.
Inked line art with flat colour, comic panel framing.
- Prompt
- a robed figure alone on a mountain summit, raising a staff of light into a storm, cinematic realism, volumetric god rays, cold blue and gold
- Avoid
- flat lighting, extra limbs, text, watermark
- Steps
30, guidance7.0, size1024 × 1024
Placeholder prompts for the three images above, written to show the shape rather than to be reproduced exactly. They will be replaced with the real settings, model by model.
Where it shines
- Illustration, concept art, key art and stylised character work.
- Iterating on your own sketch through image-to-image.
- Local edits: change the season, change the light, remove the thing in the corner.
- Batches of variations while you make coffee.
Where it will not
Hands, crowds and long strings of legible text remain the classic weak points. Precise compositional control — this object exactly here, at exactly this angle — is still easier to achieve by painting a rough guide and running image-to-image over it.
Video Experimental
Short clips from a sentence, or motion added to a single still you already like. The most demanding panel in the studio by a wide margin, and the one where expectations need managing most carefully.
On your shelf
Text-to-video and image-to-video models, plus a separate text encoder they share. This is the largest download in the app, and the model files are measured in gigabytes rather than hundreds of megabytes.
Where it shines
- A few seconds of ambient motion behind a title card.
- Giving a still illustration a gentle drift — cloth, smoke, water, hair.
- Mood tests and animatics, where roughness is acceptable.
Where it will not
Clips are short, resolution is modest, and a render can take many minutes. Character identity wanders between frames, cuts are impossible, and dialogue is out of scope entirely. Plan on several attempts per usable shot.
We label this experimental deliberately. On-device video is at the stage image generation was at a few years ago: real, occasionally magic, and not yet something to promise a client on a deadline.
Music Early
Sound from a description, built on the Stable Audio family: a loop, a bed, a sting, a texture. Useful precisely when you need something under a cut and do not want to spend an hour in a library.
On your shelf
Two directions from the same family — one tuned for sound effects and short textures, one for musical material — plus a larger model when you want more body.
Where it shines
- Sound effects and foley-adjacent textures: rain, a door, a distant market.
- Loopable ambient beds for a scene or a menu screen.
- Short stings and transitions, generated in a handful of takes.
Where it will not
Do not expect a song. Structure over a full arrangement is where the current generation of open audio models stops: verses do not develop, choruses do not return, and vocals are not on the table. Clip length is limited, and the further you push it the more the material wanders.
Speech to text Ready
Drop in a recording and get text you can edit. An interview, a voice memo, a lecture, a client call you were allowed to record — transcribed on your own machine, which for this particular task matters more than for any other.
On your shelf
Whisper-family models in several sizes. The small ones are quick and adequate; the larger ones are noticeably better with accents and background noise.
Where it shines
- Clean single-speaker recordings, in many languages.
- Confidential audio that must not be uploaded to a service.
- Getting a rough transcript fast, then cleaning it in the Text panel.
Where it will not
Overlapping speakers, heavy accents, poor microphones and dense jargon all cost accuracy. Assume a proofreading pass on anything you intend to publish.
Text to speech Ready
Give a script a voice. Read a paragraph aloud in a stock voice, describe the voice you want in plain words, or keep one voice consistent across a series of takes.
On your shelf
Speech models in base, custom-voice and voice-description variants, in two sizes. The description variant is the interesting one: you write what the voice should sound like instead of hunting through a list.
Where it shines
- Narration for a walkthrough, a trailer or a prototype.
- Placeholder voice lines while a game or animation is still moving.
- Hearing your own copy read back, which catches clumsy sentences fast.
Where it will not
Prosody drifts over very long passages, so cut scripts into paragraphs. Fine emotional direction is still coarse. And an obvious point worth stating plainly: cloning a real person's voice without their consent is not a use we support.
Vision Ready
The one panel that reads instead of writes. Hand it an image and ask what is in it: a caption, the text on a sign, the boxes around the objects, the outline of a subject you want to cut out.
On your shelf
A stack of compact vision models rather than one big one: OCR, classification, detection, segmentation, face landmarks and attributes. They are small, they load fast, and several of them ship as part of the app.
Where it shines
- Captioning a folder of images so you can search it later.
- Pulling text out of a screenshot, a scan or a photographed page.
- Finding and outlining a subject before an image-to-image pass.
- Sorting a shoot by what is actually in each frame.
Where it will not
Handwriting, stylised type and low-resolution scans push OCR accuracy down. Detection works from a fixed vocabulary of classes, so unusual subjects come back as the nearest guess. Face attributes are approximations, and we would rather you did not treat them as facts about a person.