Generator

The generator is the main screen of the service. Images, video and sound are made here. Your works are on the left, the canvas and the prompt field in the middle, the settings of the chosen model on the right.

Generator, Images tab

Three tabs at the top

Images · Video · Audio — three different sets of models and three different sets of settings. Switching the tab changes the whole right-hand column.

The generator remembers the tab between visits. If you opened it and saw Video instead of images, that is not a bug — you were working with video last time.

Until something is created, an empty placeholder sits where the canvas is — that is not a loading error.

How to write prompts

The prompt field carries a hint: "What do you want to create? Write a prompt in Russian or English, or click the dice for a random prompt".

Plain language works almost everywhere. The vast majority of models are marked as accepting prompt translation, so you rarely need to translate anything yourself.

What is actually worth writing

The model does not know what you mean. "A beautiful girl" is not a prompt, it is a topic: millions of different pictures fit it, and you will get a random one. Concrete words give it something to hold on to, and usually four things are enough:

  1. Who or what is in the frame — "a red fox", "a wooden house", "a woman in a knitted sweater".
  2. Where and when — "in a snowy forest", "in a kitchen in the morning", "against a neon sign".
  3. How it is shot or drawn — "photorealistic", "watercolour", "a film still", "a flat icon".
  4. Light and mood — "soft morning light", "backlit sunset", "a warm cosy evening".

Example. Instead of "a winter house" — "a cosy wooden cabin with a red roof, a fir tree beside it, snow on the ground, warm light in the window, soft winter evening sky, storybook illustration". The second phrase gives the model everything it needs so it does not have to guess.

Don't put negations in the main prompt. "No people", "not cartoonish" often work backwards: the model sees the words "people" and "cartoonish". There is a separate "Negative" field for that — write there what should not be in the picture.

Three buttons that write for you

  • The dice next to the field inserts a random ready-made prompt. A great way to see what a "good" prompt even looks like and take it apart.
  • The sparkles ("Improve prompt") — this is the service's own improvement. It fires on click and works with any model: the service takes your short phrase and rewrites it into a detailed one. A notification "Improving the prompt" appears, and the updated text lands in the field — from there you can edit it by hand.
  • The microphone — dictate the prompt out loud. The shortcut is Ctrl + Shift + M (⌘ + Shift + M on a Mac); it stops the recording too.

Tooltips of the buttons near the prompt field

While recording, the button turns red and shows an equaliser reacting to your voice, with a countdown beside it: the recording is capped at one minute. After you stop, the text is recognised, lands in the field, and the field flashes.

Recording in progress: the equaliser and the countdown to one minute

Voice input is available on paid plans and spends coins like an ordinary generation. If nothing could be recognised, the coins come back.

If the recording does not start at all, it is almost always the browser permission. The first time it asks by itself, and that question needs a "yes"; if you have already refused, use the lock icon left of the address → site settings → microphone.

The browser asking for microphone permission

The button has its own hint for every case — access denied, no microphone found, microphone busy with another program. And "we heard nothing" is not an error at all: the button returns to its normal state by itself after a couple of seconds.

There is a fourth route: upload a picture and press the magnifier button on its preview — the service writes the prompt from the picture itself. Details in Special features.

Don't confuse it with "Prompt extend". The sparkles button is the service's improvement, on demand. The "Prompt extend" toggle in the "Additional" block is a setting of the model itself: it exists not on every model, and where it does it may be on by default. Then the model reworks the wording on its own side, and you will not see that in the prompt field.

If you know exactly what you want and the result keeps drifting away from it — check that toggle and try switching it off.

How much text fits

Next to the price you can see how many characters still fit. The limits differ between models, and some models have none at all — then the counter says so. Both states are normal.

Prompt length counts more than your own words: the prompts of selected styles and the trigger words of selected LoRAs are added to it. For video and sound only your own text is counted.

Choosing a model

Under the tabs there is a model search and coloured category tags: NEW, HIT, ART, PHOTO, ANIME, 3D, DESIGN, RU TEXT and others, plus a "More …" button that reveals the rest.

Tags work as a filter. Within one group they act as "or", different groups combine as "and".

The full list of model tags

Tags come in two kinds, and the colour shows which: coloured ones are categories (what the model does: photo, anime, design, editing), grey ones are the provider and the families (SDXL, FLUX, Pony, Recraft and dozens more). So you can filter models by family right in the generator.

  • NEW — the model appeared recently; the tag drops off by itself after about a month.
  • HIT — the model is currently among the most used; the set changes on its own.

Then comes the model carousel. A card shows the price per generation, the tags, and — for models above your plan — a plan badge.

The coloured dot to the left of the model name shows how busy it is right now: green — good availability, yellow — moderate, red — poor. An inconspicuous detail that saves time: a red dot means a longer wait, and it is wiser to take a neighbouring model of the same family — the result will be similar.

Another hint appears above the model search when the selected model is missing something: "The selected model requires a 'Custom Image'" or "…requires a 'Custom Image' and a mask". It comes before you launch, not after a refusal — and it settles half of the "why won't it generate" questions.

The plan badge is shown only if the model is above your plan. A missing badge is good news: the model is available to you.

The list icon next to the "Model" heading opens a panel with all models in a grid — with its own search, the same filters and bookmark stars.

Model picker panel

The carousel shows two or three models, the panel shows a dozen at once. When picking among unfamiliar ones, open the panel: comparing prices and tags there is much faster. And to understand how models differ in substance, look into Model catalogs.

Why models have different settings

The most common question sounds like this: "this model has no Negative field, where did it go?"

The answer: settings are not universal. Some blocks are switched on by the model itself, others come from its own schema. If a block is missing, the model does not support it — the interface is not broken.

This is the norm, not the exception: roughly every third model has no creativity mode, every fourth has no seed field, every fourth has no negative prompt.

A hidden block is not cosmetics, it is a ban. The server rejects a generation if you send a value for a setting the model does not have. So "this model does not have that setting" is precisely true: it will not be ignored, it will be refused. And the other way round — if the block is visible, it definitely works.

Tip. Don't hunt for a setting — change the model: the set of blocks changes the moment you pick a new one.

Layers: how the blocks of the right panel differ

Blocks in the right panel come in two kinds, and the difference matters.

  • A plain block — a title, a hint and content. It cannot be collapsed or switched off: that is how the model, creativity mode, resolution, duration and FPS work.
  • A layer — the same plus a stack icon on the left and a chevron on the right. That is your own image, the schema's media fields, "Additional", styles, LoRAs, and video frames and references.

A layer has two independent states: on or off (the stack icon) and expanded or collapsed (the chevron).

A switched-off layer does not take part in the generation but stays in place — it does not disappear, it fades. That is fundamentally different from "the model has no such block": there it truly does not exist, here you switched it off yourself.

This is exactly what explains the error "The 'Custom Image' layer is switched off": the picture is loaded, but the layer is dimmed, and the model won't get it.

Dragging is smarter than it looks: drag a work onto a collapsed or switched-off layer and it will expand and switch itself on to accept the file. No need to unfold the block first — just bring the picture to it.

A layer is a switch, not a delete: the settings inside are preserved. Built a frame with a reference and want to see "how it looks without it" — don't remove the picture, dim the layer and run again.

If a block has no stack icon, it cannot be switched off, and looking for a toggle is pointless. Switching layers off is available on paid plans: on the free one, clicking the icon opens the plans window.

Resolution

Resolution is set in two steps: first a quality preset — every model has its own set, from 0.5K up to 4K — then proportions with a slider. While you drag the slider, the exact size in pixels shows above it. Switching the preset keeps the chosen proportions, and the price is recalculated immediately next to it.

The Resolution block: presets, the proportions slider and the exact size

Video has its own presets instead — from 480p up to 4K, a different set on every model; the preset row only appears at all if a model has more than one. Duration appears there too.

Choose the shape of the frame first, then raise the quality. When you switch presets the service looks for the closest aspect ratio in the new one, so a vertical frame stays vertical — the shape will not slip.

The set of sizes changes when a custom image is loaded. On some models the list narrows to a fixed one, and it looks as if the options vanished. Remove the picture and the list comes back. Hence the trick: if the proportions you need are missing — take the custom image off, choose the size, then put the picture back.

And on video models that take the size from the source, the proportions slider goes dark altogether: the aspect ratio is read from the file. The quality presets keep working.

Next to it sits the number of images per run. How many options there are is decided by the model; which of them are available to you, by your plan. Unavailable ones are not hidden but shown inactive, and clicking such a number opens the plans window.

The price is multiplied by the count: four pictures cost four times as much. Try an unfamiliar model on one, and raise the count once the prompt has settled — that way you pay for exploring variants, not for exploring wordings.

Your own image

A file can be dragged from disk or straight from your own feed on the left — no need to download and re-upload. The drop zone lights up with a dashed border when what you are dragging fits; it ignores the wrong file type, and that is not an error.

The number of slots is set by the model and written right above the frame: "Available to upload: …". When the slots are full, the drop zone disappears completely — it does not go grey, it goes away. Resolution adapts to the first file automatically.

Custom image loaded

The preview itself carries four buttons in its corners, each with a hint: crop, remove image, extract prompt from image and edit image — the last one opens the uploaded file in the editor.

The "Crop image" hint on the preview

The "Edit image" hint on the preview

Once the file is in place, "Draw mask", "Remove background" and "Modification strength" appear. The last one decides how far the model departs from the source: the scale has labelled notches — Minimal / Medium / Maximal. To repaint an object while keeping the angle and background, stay closer to minimal; for a new scene inspired by the source, closer to maximal. The slider is not on every model, and its range is set by the model.

If the picture is not yet saved among your works, it is saved automatically before the background is removed — leaving you with two works: the source and the result. Those are not duplicates.

On models that require a mask, the drawing window opens by itself — once, right after the upload. Close it without saving and it will not come back on its own; after that only the "Draw mask" button reopens it.

Uploading a personal photo — switch private mode on yourself, with the toggle in the right panel: it does not come on by itself. Then the work will not go into the public feed. Private mode is available on a paid plan.

Four buttons on the preview itself

WhereWhat it is
top left"Crop image" — crop again, after uploading
top right"Remove image"
bottom left"Extract prompt from image" — the service writes the prompt for you
bottom right"Edit image" — opens the editor right from here

The top-left button changes meaning: if the picture was uploaded as a file, it crops; if you drew it, the same button takes you back into the drawing window to finish it.

To swap the reference you don't need the bin — just drop another work onto the preview. But note: a drop onto an occupied slot replaces the file rather than adding a second one, and the painted mask is wiped along with the old picture — it was tied to that image. The mask has to be drawn again.

Cropping is not "trimming the file" but a required step. The window is named accordingly: "Set the visible area to use the image for generation". On some models it pops up by itself right after the upload.

And here is the main thing: until the visible area is set, the generation will not start. Close the window without pressing "Save" and the block shows an explanation with a "Set visible area" button instead of settings. This is the single most common cause of "the Create button doesn't work": the person is sure they just closed a window they did not need.

The frame is tied to the chosen resolution rather than free: change the size and the frame changes with it.

If the block shows an explanation instead of an upload area, there are only four reasons: your plan does not allow it (your plan will be named in the text), the model does not allow it, both at once — or exactly the visible area is not set. The first three are refusals; the last is an unfinished step.

Draw instead of uploading

Next to the upload button there is a brush. It opens a window "Draw your own image" — a clean canvas where you can sketch the composition by hand. Below the canvas: brush size, a colour field with an eyedropper, a brush / eraser toggle, clear, and the resolution of the future picture.

Drawing your own image

Example. On the left of the shot is a sketch drawn with a mouse in a minute: a box for the house, a triangle for the roof, a fir tree, a ground line. Here is what came out of it:

Result from the sketch

The composition survived intact: house in the centre, fir tree on the left, roof of the same shape. Everything else — light, snow, wood grain, smoke from the chimney — the model filled in.

A sketch alone is not enough — the model matters. On a photorealistic model a line sketch stays a line sketch, even with "Modification strength" turned up. This worked on a model tagged EDITOR, with the prompt written as an instruction: "turn this line sketch into a warm storybook illustration…". If the result stubbornly resembles your drawing, change the model, not the slider.

Mask: show exactly what to change

If the model supports it, "Draw mask" appears under the preview. The window says that the painted area will be changed during generation.

Mask editor

There are three tools: brush size, a brush / eraser toggle and invert.

Invert saves the most time. Want to keep the character and replace everything around them — paint over the character (that is quick) and hit invert. No need to paint the whole background by hand.

On some models the button is replaced by "Mask is not available for the selected model", and on erasing models — the opposite: "A mask is required for the selected model". That is a property of the model, not a fault.

If the file did not upload

What is wrongWhat the service shows
wrong format"Unsupported file format. Use jpeg, png or webp"
file too heavy"The file is too large. Maximum size is 20 MB"
file too light"The file is too small and cannot be uploaded"
file unreadable"Could not read the file. Try another image…"
too many files at once"You selected too many files"

Styles, LoRAs and additional settings

Additional settings

Below come the blocks. Your own image, styles, LoRAs and "Additional" can be collapsed and switched off — they carry the stack icon. Private mode, "Noise" and the image count are plain blocks: you can only fill them in. Prompt extend is not a block of its own but a toggle inside "Additional", and not on every model.

You can pick no more than three styles and three LoRAs. The trigger words of selected LoRAs are appended to the prompt automatically — you don't type them by hand, but they do eat into the prompt length.

A selected style or LoRA has a toggle and a cross, and they are different things: the toggle switches it off temporarily without losing it from the set, the cross removes it altogether. Handy for comparing "with and without" instead of hunting for the style in the ribbon again.

The five icons above the styles ribbon are not a filter but a change of "sitter". Cat, apple, tractor, man, woman: the previews of all styles are redrawn on the chosen subject. An inconspicuous control that decides a lot: element of the block — a style that looks great on a cat can fall apart on a portrait.

Planning a portrait — switch the sitter to a person first, and only then pick a style.

A style is not a black box: its card shows the prompt itself, and you can copy it (by pressing the heading, not by selecting the text). The style's "Avoid" is there too, if it has one.

Hence the best trick of the whole topic: open the prompt of a style you like and take its wording into your own. It works even where the style itself is above your plan.

The order of styles matters. They do not blend; they stack in turn, each one wrapping the result of the previous. And your text may end up inside the style's template rather than before it — which is why short wordings work better with styles. Write the substance; the style supplies the framing.

Three ways to get "nothing happens" with a LoRA, and all three are silent:

  1. three are already chosen — a click on the fourth simply does not fire, with no message at all. Remove one first;
  2. the LoRA is above your plan — the card is covered by a gradient with a large padlock, and a click opens the plans window instead of selecting it;
  3. the whole LoRA layer is switched off — the LoRAs are chosen and visible, yet they never reach the generation. The same goes for individual toggles inside the panel.

What happens to LoRAs when you change the model is the commonest "bug" that is not a bug: incompatible ones disappear silently, with no message; the strength you set survives only if it falls inside the new model's range, otherwise it reverts to the author's default; a LoRA switched off with the toggle stays off; anything beyond three is dropped. Compatibility is not decided by the family name: if a LoRA is visible in the list under your model, it fits.

Video has no LoRAs at all — not a single video model accepts them.

The list of LoRAs is assembled for the selected model, and each one's weight is already set to the value its author recommends. Don't turn the strength blindly and don't stack three at once — they almost always fight each other; switch them on one at a time with the toggle. An empty LoRA list is not a connection failure but a property of the model: it does not support them. More in Model catalogs.

"Noise" does not guarantee a repeat if prompt extension is on. The setting's own description says as much: it affects reproducibility. To repeat a work exactly, set "Noise" and switch prompt extension off in "Additional". The seed alone is not enough.

Three of the common ones are worth knowing. "Prompt adherence" — how strictly to follow the text: the slider has labelled notches from "Low" to "Strict". "Noise" is the seed, the number that lets you repeat a frame; a green dice beside it means the value is random. And "Grain" is about texture, not focus: it makes the picture more textured, but it will not rescue a blurry frame.

The set of additional settings comes with the model itself, and every model has its own. SDXL models bring "Prompt Adherence", which newer models do not have; design models bring a colour palette and a background colour; some models bring "Prompt extend". "Creativity mode" is not from "Additional" at all — it is a separate block of the right panel, right under the model. It has three steps — Sketch, Draft, Masterpiece — and the captions at the ends of the slider say plainly what you pay with: "Faster" on the left, "Higher quality" on the right.

Creativity mode: Sketch, Draft, Masterpiece

The practical takeaway: changed the model — scroll through "Additional" again. A setting you were used to may be gone, and a new one may have appeared.

Your own feed on the left and dragging

The arrow button at the left edge opens a drawer with your works: search, grid density and media type switches. From there a work can be dragged straight into the "your image" slot.

Left drawer with your works

The shot also shows a deleted work — red, with a bin and a restore arrow. Until you leave the page it can be brought back with one click. More in Your works.

Hover a card in the drawer and an enlarged preview appears beside it, with the resolution, the date and the first line of the prompt on top. It is the fastest way to find the right work among a hundred thumbnails — no need to open a card and come back. Video plays silently in the preview and audio gets a waveform player, so a track can be listened to without opening it.

The drawer has its own search — it looks through the words of your prompt and takes the first ten words into account. Search by one or two rare words: a long phrase gets truncated anyway. It also has its own density slider: a single column turns the drawer into a large preview feed, which is the most convenient way to pick.

The right panel can be collapsed too — then only the canvas and the prompt field remain.

Collapsed right panel

On a phone this works differently

The drawers open with a swipe from the edge of the screen — an invisible trap strip runs along the very edge. A short, brisk movement is enough; there is no need to drag all the way.

Only one drawer is open at a time. Open your works and the settings close by themselves — and the tab of the second drawer is not shown at all. Hence "the settings panel disappeared on my phone": it did not, just close the drawer with your works.

Two lesser-known tricks. First: the tab can be dragged up and down with a finger, and it stays where you left it — separately for the left and the right drawer, and it survives a reload. That saves you when the tab covers the button you need. Second: at the bottom of an open drawer there is a wide "Close gallery" button — if swipes are not your thing, close it with that.

Also on a phone the improve-prompt button moves out of the input field into the bottom bar, the Ctrl + Enter hint is not shown, and the create button itself becomes round with an upward arrow.

Launching and waiting

The launch shortcut is Ctrl + Enter (⌘ + Enter on Mac) and works anywhere on the page. On a phone the hint is not shown and the button is round, with an upward arrow.

Generation in progress

While the work is being made, a notification hangs at the top and a placeholder with a spinner sits in the feed on the left. You don't need to leave the page, but you don't need to watch it either: the result appears by itself.

Works run in a queue, so time may pass between the press and the result — that is not a freeze.

Read the error text straight away. Finished rows disappear from the panel after about a minute, and then there is nowhere to look up the cause — hence the eternal "something red flashed and I missed it". The work itself does not go anywhere: a failed attempt stays in the feed as a cross in a circle.

A generation also has a time limit: if the network never answers, the job closes itself with "The generation time has expired" — and the coins come back. Slow models wait longer, and that is normal.

If several pictures were ordered, a row of thumbnails appears under the big one — for switching between variants of the same batch. With a single picture there is no row. A finished result can be dragged straight from the canvas into the settings fields, just like from the left drawer. That works for images and sound; video cannot be dragged.

What you can do with a finished work

As soon as the work is ready, a row of buttons appears above it — and this is the most useful place in the generator. The buttons form three meaningful groups.

A finished result and the action buttons

The buttons are icons only, but each has a tooltip on hover — and the tooltip is what explains them. Here they are in order, left to right.

Send to "Custom Image"

Puts the finished result into the source slot, so the next generation starts from it. That is how a chain is built: generate, send to the source, rewrite the prompt, get a development of the same picture.

The "Send to Custom Image" hint

The button is smarter than it looks, because the source size has to suit the model. If the result matches the current size, the picture is simply placed in. If the model has that size among its own, the resolution switches by itself. And if there is no suitable one, the cropper opens — otherwise the model would not take the file. The canvas is cleared, and the "Custom Image" layer switches on and unfolds on its own: a clean field for the next step.

Create a similar image

Runs the very same prompt with a random "Noise" value — and makes exactly one picture, however many were set in the settings. It is the cheapest way to get another take on the same idea: nothing needs to be changed.

The "Create a similar image" hint

That is why the result differs every time: the seed is reset to random. If you need an exact repeat rather than another take, set "Noise" by hand in the additional settings.

Animate (image to video)

Switches the generator to the "Video" tab, puts your picture in as the first frame and unfolds the frames block by itself. What is left is to pick a video model and write what should happen in the shot.

The "Animate (image to video)" hint

The prompt is not carried over: video has its own logic of description — there you write about movement, not about the content of the frame.

Discuss with AI in chat

Uploads the picture as an attachment and opens the AI chat with it. While the file is uploading, a spinner runs in place of the icon — that is a couple of seconds, not a freeze. If the upload fails, a notification says so and nothing is lost.

The "Discuss with AI in chat" hint

Useful when you need to take your own work apart in words: ask it to describe the composition, suggest what to improve, or write a prompt for the next generation from it.

Upscale resolution

Enlarges a finished picture without rebuilding it. The button opens a menu with two items — x2 and x4.

The "Upscale resolution" hint

The upscale menu: x2 and x4

The prompt plays no part here at all: upscaling works with pixels, not with a description. The result arrives as a separate work.

Upscale resolution x2 + prompt

The second upscale is always x2, with no choice of factor, but it paints in detail from your text. It takes the prompt and "Negative" from the fields as they are right now, not the ones the picture was made with.

The "Upscale resolution x2 + prompt" hint

With an empty prompt field the button will not fire and shows an error about the missing prompt. That is the whole difference between the two: one simply enlarges, the other enlarges and paints in detail from the description.

Outpaint the image

Extends the frame beyond its borders: the service paints in what was never in the shot. The button opens a menu of four sides — top, left, right, bottom.

The "Outpaint the image" hint

The outpaint menu: top, left, right, bottom

Sides are picked one at a time, and each pass produces a new work. That is how a tight composition turns into a wide shot: a few outpaints in a row and a portrait becomes a scene. One pass adds roughly a quarter of the side, but no more than a fixed maximum — which is why a wide shot is built in several steps rather than one.

Outpainting has two limits, and both arrive as a message. The same work cannot be outpainted forever: after a few passes the service says the limit is used up. And it will not outpaint a picture that is too large — it says that outright too.

Hence the order of operations: outpaint first, upscale afterwards. An upscaled frame will almost certainly hit the size limit.

Background remove

The only button with no menu: press it and the work goes into the queue straight away. What comes back is the picture with a transparent background.

The "Background remove" hint

Puts the link to the work's page into the clipboard — the very page anyone following it will see. It is a link to the page, not to the image file: next to the work they will find its prompt, model and settings.

The "Copy Link" hint

Download

For pictures the button opens a menu of formats — WebP, PNG and JPG; vector models add SVG, because the source itself is vector there.

The "Download" hint

The download menu: WebP, PNG, JPG

Video and sound have no menu at all — the file downloads at once, in one press: a clip as MP4, a track in its own format. Raster formats do not apply to them, so no choice is offered.

On the free plan downloading is available in WebP only, PNG and JPG come with paid plans. And a separate case: a work marked 18+ will not download while adult content is switched off in profile settings — the settings open instead of the download.

What is available on the free plan

ActionFreeAny paid plan
Create a similar imageyesyes
Face restorationyesyes
Copy link, downloadyesyes
Send to "Custom Image"noyes
Upscale and upscale with promptnoyes
Upscale together with face restorationnoyes
Outpaint beyond the framenoyes
Background removenoyes

Face restoration is the only enhancement available on the free plan. Upscaling itself is not, and neither is "upscale together with face restoration": that is a separate action with its own restriction.

A dimmed button and a highlighted one mean different things. Dimmed means there is nothing to work with: a generation is running, or there is no finished result. Highlighted means the feature exists but is not part of your plan. A dimmed button means something else: there is nothing to work with — a generation is running or there is no finished result. Highlighted means the plan is missing; pale means the occasion is.

While the buttons load, identical placeholders sit in their place: the panel does not jump, but it does not work either — that is not a freeze. And if there is no panel at all, there is no result: every button hides itself on an empty canvas.

Video and sound

The "Video" and "Audio" tabs work the same way but with different settings.

Generator, Video tab

Video brings duration, frames per second and resolution — from 480p up to 4K on some models. Video is charged per second, so duration multiplies the price directly.

Video: 4K resolution, duration and keyframes

The list of durations depends on the chosen size. On some models 4K offers one set of durations and 1080p another, and when you change the size an unsupported duration moves itself to the nearest allowed one. It looks like the interface acting on its own; in fact it is protection from a refusal.

Hence the tip: if the duration you need is missing, try a different size. On a number of models long clips are only available at a lower resolution.

What goes where in video

On video models files go into different slots rather than one, and this is the main confusion of the section:

SlotWhat goes in
Framesliterally frames of your clip: the first, the last
References (images)a sample of what it should look like: style, character
References (video)movement or manner to be repeated
References (audio)sound as a sample
Audiothe track that will be part of the clip itself
Source videothe clip you are editing (on editing models)

Only the slots a model supports are shown — usually two or three, not all of them. Each is a layer: it collapses, switches off and accepts drag-and-drop from the left drawer.

The difference between "Frames" and "References (images)" is the commonest mistake. A frame is literally a frame of the future clip — the model will start (or end) with it. A reference is only a sample to imitate; it will not appear in the clip.

Uploaded audio — check the frames per second: some models have their own set of rates for a voiced clip, and the value may have jumped by itself. And remember that on some models the duration is taken from the length of the track — a one-minute audio gives a one-minute clip and a one-minute price. Trimming the track beforehand is cheaper than generating spare seconds.

"Extract a frame" takes neither the first nor the "best" frame, but the one currently paused in the player. So the order is: pause the clip at the moment you want — and only then press the button. The frame is taken at the video's full resolution, goes straight into "Custom Image" replacing the previous reference, and the generator switches to the "Images" tab. If the frame could not be captured, the clip's cover is used instead.

A finished video has only three actions: extract a frame, copy the link and download. There is no upscaling, no outpainting and no "create similar" for video — don't look for them. The only way to refine a clip: extract a frame, fix it as a picture and animate it again.

Generator, Audio tab

Audio model tags split them into clear groups: SPEECH (voicing text), VOICE-CLONE (voice cloning), MUSIC (music and songs), SFX (sound effects). Voicing is charged per character, not per second — which is why it has no duration slider.

A finished track has two actions: copy the link and download. A track cannot be animated, sent to the chat or redone with one button — only by a new generation.

Both topics are covered in detail in Special model features.

Resetting the settings

The "Reset" button next to "Create" does more than expected: it clears both the prompt and "Avoid", returns the model and every setting to their defaults, and removes the previous result from the canvas. There is no warning beforehand — you press it and it is done.

Your works are not deleted — they stay in the feed on the left and in your profile; only the form is reset. But if the prompt matters, copy it first: it cannot be brought back afterwards. Or, instead of resetting, simply open any of your works from the drawer — its settings land on top of the current ones.

What requires a paid plan

Some of the generator's features work on paid plans. If one of them is not open to you right now, the service says so and names it: generating from your own image and a mask, private mode, the "Masterpiece" creativity mode, premium styles, voice input, prompt improvement, extracting a prompt from a picture, extending the frame beyond its borders, background removal, upscaling together with face restoration, and models above your plan.

No coins are charged for such a refusal. What exactly each plan includes is in the plans window — see Plans and payment.

Three of the messages are written in technical language and deserve a translation: img2img means generating from your own image, outpaint means extending the frame beyond its borders, and prompt generation from an image is that magnifier button which writes a prompt from a picture.

"Masterpiece" has an exception: on a few models it is open even on the free plan. The rule is simple — if the mode button is active, this model is one of them.

If the generation won't start

What happenedWhat to do
the prompt is shorter than five characterswrite more: the button stays dim while the text is too short
the prompt is empty and the model requires onewrite something
the visible area is not set on the uploaded picturepress "Set visible area" and save
the model requires an image or a maskadd what is missing
the "Custom Image" layer is switched offswitch the layer on with the stack icon — the picture is loaded but not being sent
a required field in "Additional" is emptylook for the field with an asterisk — it is highlighted
not enough coinsclaim the daily ones or top up — Coins and limits
the simultaneous generation limit is reachedwait for one of the running ones to finish
another generation is runningthe button does not dim, it spins — that is not a refusal
no prompt entered"Upscale with prompt" does not work with an empty field
the feature is not available on your plansee the section above and Plans and payment

There is one more case: a temporary block for repeated forbidden requests. Send something the service refuses several times in a row and generation closes for a while — the message says how long is left. The block lifts by itself; there is no need to write to support, just rephrase the request.

On the free plan an advertisement is shown in place of the empty canvas. It can be closed, but not instantly: a short countdown runs next to the cross. After your first generation the canvas is taken by the work, and any paid plan removes the advertising entirely — both here and in the feed.

Where to go next

Edit on GitHub