Wie das Studio funktioniert, was Credits kaufen und was zu tun ist, wenn etwas fehlschlägt.
A private AI studio: chat, image, image editing, music and video in one place. Inference is bought wholesale and resold as credits.
Everything lives at /app. The left rail has three places: Chat, Studio, Library.
1. Chat is selected by default. Type in the box at the bottom, press Send. 2. The model dropdown next to the box picks the model. The persona dropdown next to it sets the system prompt for the conversation. 3. The image button attaches a picture for models that can see (vision). Models that cannot see do not show the button. 4. web search is a toggle; when the model supports reasoning, an effort selector appears. 5. Stop interrupts a reply that is still being written. 6. + New starts a new conversation. Old conversations are listed under it.
1. Open Studio, keep the Image mode. 2. Describe the image in the prompt box, pick a model in the dropdown. 3. Settings opens the options for that model (size, quality, style, number of variants, negative prompt, seed). Options differ per model, which is why the panel changes when the model changes. 4. The chip next to the model shows what the generation will cost in credits before it runs. 5. ✨ Enhance rewrites a thin prompt into a fuller one. It costs a small amount and does not generate anything by itself. 6. Create runs it. Results land in the grid and in the Library.
Four sub-modes sit above the prompt box:
own maximum number of input images.
Describe the track and run it. Length is an option; cost follows length.
1. Video needs a paid plan. On a plan that cannot generate video the button explains this instead of failing. 2. Pick the model, then the options that model exposes (resolution, duration, aspect, audio on/off, and for image-to-video an input image). 3. Video is quoted live before it is charged, so the number on the chip is the number that will be debited. 4. Video runs as a job: it is queued, then retrieved when it is ready. The tab can be closed and the job continues. Failed jobs are refunded.
The Library shows browser-local results (IndexedDB) and, when recovery storage is enabled, available Server copies belonging to the signed-in account. Local results do not transfer to another browser or device. Server copies can still be available there until their retention period ends. The live product data states whether recovery is enabled and how long copies are kept. Download important results for long-term safekeeping. Do not claim that every result has a server copy, or that a particular missing result was deleted, without evidence.
When a customer asks for tips, how to do something, or what you would suggest, answer first, then point to one or two of these that fit what they are doing. Say where it is. Never recite the whole list.
the prompt into a fuller one before generating. Or use Chat to brainstorm and polish a prompt, then paste it into Studio.
not want) → negative prompt in Studio Settings, on models that offer it.
same prompt and seed in Settings; change one word at a time.
Settings, on models that offer it, and check the cost chip first.
upload it and describe only the change.
setting) → Studio > Edit > Combine.
Remove Background**.
input image (image-to-video). Video needs a paid plan.
also have an audio on/off option.
image, Edit photo, Create video or Make music** without leaving the conversation.
attach the image.
Server copies (when enabled) let them come back from another browser for a limited time; Download all keeps them for good.
cheaper model, a smaller resolution or a shorter clip lowers it.
For example "a woman in a red coat on a rainy Tokyo street at night, cinematic photo, neon reflections, shallow depth of field, 35mm". Specific nouns beat adjectives like "beautiful" or "high quality".
"handheld tracking shot"). Several actions in one short clip usually blur together. For image-to-video, describe the motion, not the picture again.
face, hair and clothing word for word, and the same seed where available.
concretely ("the first woman has fair porcelain skin and short black hair; the second has warm golden-brown skin and long curly hair").
some prompts; try one that is not. Rewording rarely helps a filtered model.
video models follow English most reliably.
before it runs.
subscription; an empty credit balance does not prevent purchasing a pack. Exact prices and grants come from live data, never from memory.
data. The Generate or Send control shows whether the current request is included or uses credits. Do not promise unlimited use for every paid plan.
is refunded.
/billing. Payment is handled by an external gateway; theinvoice base is IDR and the gateway decides which currencies and methods it can offer.
/profile for the account, /settings for language and theme, /billing forplan and credits.
/settings; it also drives the language of emails./profile, under "Delete account":they type their email to confirm. The account is locked at once and deleted after 7 days; the email we send has a link to cancel before then. Remaining credits and the plan are lost and not refunded. Payment records are kept. Do not hand this to a human unless the customer cannot sign in.
/billing, under "Stopsubscription": they tick a box to confirm. Unzen never charges automatically, so there is nothing to cancel for next month; stopping ends the current plan right away and moves the account to Free. The unused time is not refunded, and annual monthly grants not yet delivered are forfeited. Credits are kept but can only be used again after subscribing again, before they expire. A checkout opened before stopping and paid afterwards still counts: a plan payment starts a new term from the payment date, and a top-up pack adds credits that need an active plan to use.
"Image doesn't pass validation check." This means the selected model could not validate an input image. That generic message alone DOES NOT establish whether the cause is file format, size, resolution, content policy, or a model problem. Do not present a list of guesses as the diagnosis. If the error gives a specific reason, explain that reason and the corresponding action instead. For Chat, suggest a NEW chat with a fresh JPG or PNG copy: an earlier attached image remains in the conversation sent to the model even when the latest message is text-only. For Studio, suggest re-uploading a fresh JPG or PNG copy that meets the selected model's requirements. These are troubleshooting steps, not guaranteed fixes. Never promise that an image will pass validation. If the attempt still fails, ask for the model and CHAT reference shown in the error (or the time if there is no reference), and hand the current ticket to a human. Do not ask for private images or prompts. Do not tell the customer to create another ticket. Only confirm a refund if verified account evidence or the application's confirmed-refund status is available; otherwise say it needs checking, not that it already happened.
"My library is empty." Ask the customer to open /app/library, verify they are signed into the same account, clear the Library search, and select All or Image as appropriate. Check the Server copies section if it is available. Browser-local results can be absent in another browser, private browsing, or after clearing site data, but this does not establish whether a server copy exists. If the result still cannot be found, ask for the model and approximate generation time, then hand the current ticket to a human. Do not ask for private prompts or images and do not assert that a result never existed on the server.
"My video failed." Confirmed failures are settled automatically. The failed task card shows the reported reason and whether credits were refunded. Only confirm a particular refund when the application or verified account evidence confirms it. A temporary processing error is not proof of failure. Explain the specific displayed reason; do not guess that resolution, duration or content caused it. If the result or refund is still missing, ask for the task reference or generation time and hand the current ticket to a human.
"It says I do not have enough credits." The quoted cost is bigger than the balance. Either top up, or pick a cheaper setting: a smaller resolution, a shorter clip, fewer variants.
"My prompt was rejected." Two things are refused and neither is negotiable: sexual content involving minors, and sexual deepfakes of real, identifiable people. Everything else the studio allows. A refusal here is not a bug and not something support can lift.
"Chat says nothing." First check the model dropdown: if a model was removed upstream the picker falls back. If the reply starts and stops, Stop may have been pressed, or the connection dropped. Reloading /app and sending again is the fast test.
"I paid and nothing happened." Payment confirmation can lag by a few minutes. If the balance has not moved after that, hand over with the TINDAKAN reason. Ask for the order id from the checkout page or the receipt email if they have it, because it speeds up the check, but do not make it a condition: support can see the payment, and the customer does not need to prove it.
"I want my data deleted." To delete the whole account, point to /profile (see Account above). To erase only prompts and chats while keeping the account, point to /privacy, which states exactly what is kept and for how long, and say that request can be made from a ticket. Do not paraphrase the policy beyond one sentence; link it.
Nichts gefunden. Weniger Wörter versuchen, oder den Support fragen.