The Sandbox

ElevenLabs text sound effects use Sound Effects v3, supporting 0.5 to 30 seconds. Sound Effects v2 is no longer offered for new generations. If the v3 connection or its verified price is unavailable, try again later. Existing results and submitted jobs remain available in History; an old v2 recipe needs v3 and a fresh price before generating again.

The Sandbox is for making one thing at a time, quickly, without a project around it. Everything you

make there is kept on the page — ready to run again, download, or send into a project.

Extend your video results from Templates also appear in your library. Their recipe keeps the uploaded source, direction, continuation settings, returned seed and actual video duration. Download the result or send it to a project; open its template for another extension.

Looping video results from Templates

also appear in your library. Their recipe keeps the source image, motion instructions, motion style,

duration, resolution and audio choice. Download the video or send it to a project; open the looping

template to make another. Interrupted requests are checked using their original job ID.

Talking image results from Templates also appear here. Their saved recipe keeps the face image, speech audio, chosen resolution and returned seed; download the video or send it to a project from its result.

Recast video results from Templates also appear in your library. Their saved recipe keeps the source video, reference photos, replacement roles and generation settings. Download the result or send it to a project.

FLUX.3 Klein images

Choose FLUX.3 Klein or FLUX.3 Klein Base in Image. Write a detailed description to generate, or add up to ten PNG, JPEG or WebP pictures to edit and compose. Numbered thumbnails show the order used by the model; use the arrow buttons to reorder them. Describe the complete result and refer to the first picture, second picture, and so on.

Output settings offer square sizes of 768, 1024, 1536, 2048 or 4096 pixels, PNG/JPEG/WebP, prompt expansion and an optional seed. How many selects one to four images in a single request. Check Generate's price tooltip before submitting. Base costs more at each size. Final charges follow the supplier's actual bill plus the configured fee.

Open Layout to draw labeled objects. New objects have an output box. Reference objects can be kept, moved/resized or removed using a source box and, where needed, an output box. Drag a box to move it, use its corner to resize, or enter fractional coordinates. Insert label in prompt adds a token such as <fox> to connect your description to the box. If you remove a reference used by an object, reassign or delete that object before generating.

Open a finished result and choose its sample to preview, download or send that image to a project. Tweak restores reference order, layout, format and size. Interrupted jobs are checked by their existing request ID. Steps, guidance and negative prompts are unavailable because the current API schemas do not expose those controls.

Make something

If the connection to a submitted generation is interrupted, check History before generating again. Recover result checks the existing request. For project generations, recovered media is added to the original project's Assets; results without a project remain available in History.

1. The Sandbox opens after you sign in. You can also click Sandbox in the account menu.

2. Choose Image, Video, Sound FX, Music, 3D or

   Voice, Text or Analysis.

   

   

3. Describe what you want in the prompt box. The Model menu chooses what will make it; you can

   change it at any time, before or after you write the prompt. Hover or focus Generate to check its estimate. Once that estimate and your required inputs are ready, you can generate while other model prices keep loading.

   

   

4. Click Generate. While it works, the button reads Generating. The result appears in

   Your sandbox, above the composer.

Generate stays unavailable until there is a prompt, a recording for Grok Voice, or a picture for an image-only 3D model, and until any reference the model needs is

attached. If the job would cost more than your balance, the Sandbox says so beside the button.

OmniHuman V2 talking videos

Choose Video, then OmniHuman V2, to animate a person from one image and a speech recording. Add the image and a WAV or MP3 file. The audio determines the video length; there is no separate duration, resolution or aspect-ratio setting.

The Performance prompt is optional. Use it to guide gestures, expression, scene or camera movement. Supported prompt languages are Chinese, English, Japanese, Korean, Spanish (Mexican) and Indonesian.

Images can be JPEG, PNG, WebP, BMP, TIFF, GIF, HEIC or HEIF, up to 30 MB. Both dimensions must be 300–6000 pixels, with an aspect ratio between 0.4 and 2.5. Audio must be 4–300 seconds and up to 15 MB. Invalid files are refused; they are not silently cropped, shortened or padded.

Check the USD estimate before generating. Completed videos appear in your Sandbox library, where you can preview, download, send to a project or run the saved recipe again.

OmniHuman creates a new performance from a still image. To synchronize speech in an existing Project video while retaining its original movement, use Improve lipsync instead.

Mirelo SFX 2.0

In Sound FX, choose Mirelo SFX 2.0 Text to describe a sound, or Mirelo SFX 2.0 Video to Audio to create a soundtrack from an uploaded MP4. In Video, choose Mirelo SFX 2.0 Video Soundtrack to receive the video with its new soundtrack. Video guidance is optional.

Open Output settings to choose duration, 1-4 samples, sounds to exclude and an optional seed. Text supports 0.1-600 seconds. Video supports 1-600 seconds, limited to the source length, and requires a readable MP4 up to 512 MB. The uploaded video stays continuous across cuts. Keep dialogue is on by default; choose Sound effects only to suppress speech instead.

Audio outputs offer WAV, MP3, AAC and FLAC. MP3 and AAC also offer an optional bitrate. Video outputs are MP4. Check the estimate before Generate: it includes all samples, with the final charge based on actual billed output.

Open the finished result and use Sample to select a variation for playback, download or Send to project. Tweak restores the settings and uploaded source. Refresh checks an interrupted job using its existing request ID.

Video and 3D from a picture

Choose Wan 2.2 A14B Turbo Text to make a video from a description, or

Wan 2.2 A14B Turbo Image to animate a required first-frame picture.

Both use silent 480p output with landscape, portrait or square shapes.

The model chooses the length, so Output settings shows Model default rather

than a duration selector. Final charges follow actual supplier billing.

Play or download the saved video, or use Tweak to restore its settings and picture.

In Video, choose Wan 2.2 5B Text to animate a description, or Wan 2.2 5B Image

to start from a picture. The Image workflow requires a first frame. Choose a two to five

second length and a supported shape; output is silent 720p video. The supplier may crop the

picture to the selected shape. These workflows reserve a fixed amount per video.

In 3D, choose Trellis or Trellis 2 and add one clear picture of an object.

A description is not needed. Both return a textured GLB that you can preview and download;

they have no geometry-only or rigging choices. Trellis 2 uses its 512 resolution workflow.

The result can be ready while its final charge is still pending. Refresh results checks the

existing job and bill; it does not start another generation.

The model menu keeps Search above a scrollable alphabetical list of available models. Star a model to keep it in Favourites.

Cosmos Predict 2.5

In Video, choose Cosmos Predict 2.5 Text or Cosmos Predict 2.5 Distilled Text,

then describe the scene. Both make silent MP4 without a reference picture.

The output is fixed at 1280 by 704, 93 frames at 16 fps (about 5.8 seconds),

so Output settings shows these details instead of shape, resolution or length choices.

Check the estimate before Generate. Play or download the saved video, or choose

Tweak to restore its model and description. If its final charge is pending,

Refresh results checks the existing bill without generating another video.

Pixal3D and ReconViaGen

In 3D, Pixal3D turns one picture into a textured GLB. ReconViaGen 0.5

accepts one to four pictures of the same object; extra views can help describe its sides.

Both use fixed output settings, with no description or rigging controls to configure.

Open the saved result to inspect or download the textured mesh. Tweak restores

its model and source pictures. If its final charge is pending, Refresh results

checks the existing job and bill without generating again.

CogVideoX and multi-image Trellis

In Video, choose CogVideoX-5B Text to describe a scene or CogVideoX-5B Image

to animate one required first-frame picture. Both return silent 720 × 480 video with

a fixed 3:2 shape. The model chooses the length; Output settings shows Model default.

Play or download the saved result; Tweak restores its model, description and first picture.

A random seed is shown only when it can be stored exactly.

In 3D, choose Trellis Multi-image and add two to four pictures of the same object

from different angles. Generate stays unavailable until the second picture is attached.

This workflow returns a textured GLB, without description, surface or rigging choices.

Preview or download the mesh; Tweak restores its model and all attached pictures.

Final charges follow the supplier bill. Refresh results checks the original job and any

delayed bill without generating again.

More verified Video and 3D workflows

Choose LTX 2.3 Fast Text to make a video from a description, or LTX 2.3 Fast Image

to start from one required picture. Both include generated audio and use 1080p output.

Choose landscape or portrait and a length of 6, 8, 10, 12, 14, 16, 18 or 20 seconds.

The Image workflow uses the picture as its first frame. External audio and end-frame inputs

are unavailable in these workflows. Play or download the saved video; Tweak restores its

description, length, shape and first-frame picture.

Choose LTX 2.3 Pro Text or LTX 2.3 Pro Image for the Pro variants. They also

include generated audio at 1080p, with lengths of 6, 8 or 10 seconds.

Switching from a longer Fast video adjusts the length to the nearest supported Pro choice

and shows a notice. Pro Image requires one first-frame picture.

Choose Hunyuan3D v2 Base, Mini, Turbo or Mini-Turbo in 3D and add one clear picture

of an object. All four return GLB. In Output settings, choose Geometry only for an

untextured mesh or Textured for a mesh with materials. Textured output costs three times

the geometry price; hover or focus Generate to check the estimate before running.

These models have no description or rigging controls. Preview and download the mesh, or

use Tweak to restore the saved picture and surface choice. Earlier geometry results

keep Geometry only when you tweak them.

Choose Hunyuan3D v2 Multi-View or Multi-View Turbo to build from three pictures of

the same object. Add the Front, Back and Left views in that order. Each attached

picture is labelled; Generate becomes available after all three are ready. These models

return GLB, with Geometry only or Textured in Output settings and no description

or rigging controls. Textures cost three times the geometry price. Tweak restores all

three views in order and the selected surface.

Final charges follow actual billing; Refresh results checks the original job when its

bill is delayed.

Choose Kling O3 Pro Edit in Video, attach a clip, and describe the change. Clips are prepared

at a minimum 720-pixel short edge. The final charge uses actual billed usage and can differ from

the input-length estimate. If billing is delayed, Refresh results checks the existing job.

Choose Hunyuan 3D 3.1 Rapid in 3D to build geometry from a description of up to 200 characters.

This workflow returns GLB, without textures, source pictures or rigging controls. Preview

and download the result.

Choose Hunyuan 3D 3.1 Rapid Image to build geometry from one required picture instead.

Use a clear view of a single object with a simple background. Each edge must be 128–5,000 pixels,

and the prepared picture must be no larger than 8 MB. There is no description, texture or rigging

choice for this workflow. It returns an untextured GLB that you can preview and download.

Tweak restores the saved source picture. Rapid colour-texture and PBR workflows remain pending.

Hunyuan3D Mini and Hunyuan3D Turbo build geometry-only or textured GLB from one required picture, with no prompt or rigging controls. Choose the surface in Output settings; textures cost three times the geometry price. Check the current estimate by hovering or focusing Generate. Tweak restores the selected surface; older geometry recipes stay geometry-only. Wan 2.2 A14B Image makes silent 720p video from a prompt and required first frame, with two to five second selections. Video billing follows the supplier's final frame quantity; the estimate includes the endpoint's frame count.

Download saves the original 3D file while keeping your results open. The viewer also has

Download the 3D file. OBJ results offer separate Download materials and Download texture

buttons for files the model supplied; keep those files together when opening the object in a 3D

application. Downloads preserve the original bytes and do not generate again. Formats that the

viewer cannot display can still be downloaded for your 3D application.

More models

Open the model menu to see the category's available models immediately, in alphabetical order.

Scroll the list or search by model name or task. Favourites stay above the alphabetical list.

Click a model's star to add it to Favourites, above the list. Starring keeps the menu

open and does not select the model. Click the star again to remove it. Favourites are saved to

your account and return after closing the app or signing in on another device. Changes made

during a connection failure stay on this device and sync when the connection returns.

FLUX 2 and Seedream v5 Pro each appear once, using their current primary image workflow. Existing favourites remain saved when a catalog identity changes. Removing the star clears the older aliases too, so it stays removed when you return.

Unavailable and retired models are hidden, including from Favourites. Previously saved stars

remain on your account and reappear when their models become available. If the catalog cannot load, your

current generation models remain available; use Retry catalog to try again.

Search the Model menu for FLUX Schnell (Image), Wan 2.2 (Video), Stable Audio Small Music

(Music), Cassette Sound Effects (Sound FX), Kokoro (Voice), or Tripo P2 (3D).

Voice also offers Kokoro (British English), which reads your line with the Alice voice.

British English speech is billed in whole 1,000-character blocks; the estimate includes the

block your line starts, even for a short sentence.

Sound FX offers Stable Audio 3 SFX Base for a described sound up to two minutes; its saved

result is a WAV file you can play, download, or send to a project.

Mirelo SFX 1.6 also makes one WAV sound effect from a description. Choose a whole-second

duration from one to sixty seconds in Output settings. Its output can run slightly longer than

requested; the estimate includes the extra billed second. Final charges follow the actual bill.

Play or download your saved sound, or use Tweak to restore its prompt, model and duration.

Image also offers FLUX Dev, Sana, Recraft V2, Recraft V3 and FLUX Pro 1.1.

Recraft makes standard raster images from a prompt up to 1,000 characters. These three additions

use a description without reference images, with a fixed 1K size. Choose a shape and image count

as usual; each Recraft image is generated as its own job. Video also offers LTX Video 0.9.5 and Hunyuan Video. Hunyuan Video Image requires one first-frame picture and uses model-default length at 720p, in landscape or portrait.

These engines work from a description only. The two video engines show Model default

for length, offer landscape or portrait at 720p, and charge per video.

Wan 2.2 makes 2–10 second silent videos at 720p from a description, without a reference image.

Kokoro uses the Heart preset voice for American English. Tripo P2 accepts a description or one

picture and returns GLB; choose Textured or Geometry only in Settings.

Browse catalog includes models in all nine categories. Available models have a Use model

button. Models awaiting input, output and charging verification stay hidden until they are available.

Text supports Qwen 3.8, Nemotron 3 Nano Omni and Seed 2.0 Mini. Analysis supports Qwen 3.8 for images, Nemotron Audio Analysis, and Nemotron Multilingual Transcription.

Training currently offers browsing only.

Analyse a picture

Choose Analysis and use Model to select a Moondream, Florence 2, Perceptron Isaac, LLaVA Next or Nemotron Diffusion picture workflow. Add one PNG, JPG or WEBP picture up to 32 MB and 8192 pixels on each side. These workflows retain the full source picture.

Caption and OCR workflows need only the picture. Query workflows also need your question. Controls for unused descriptions, response lengths and analysis tasks stay hidden. Tweak picture or Tweak prompt restores the saved source and model.

Copy the result or use Download .txt for its original text. Prices are estimates; actual supplier usage determines the final charge. Refresh results checks the existing request or delayed bill without generating again.

Analyse a recording

Choose Analysis, then Transcribe speech or Ask about audio in What would you like to analyse?

Add an MP3 or WAV recording up to 10 minutes and 50 MB. Listen to the preview to check the source.

Transcription detects the language automatically and needs no prompt. Audio questions expect an English

prompt; choose Summarize this recording for a quick start, or ask your own question.

Choose a response length for audio questions, then Generate.

The price hint is a conservative estimate for a ten-minute recording. Final charges follow actual usage.

Copy or download the saved text. Tweak prompt restores the recording and model; refreshing an interrupted

request checks its original job instead of starting another generation. Switching between image and audio

clears the incompatible attachment. A transcript may contain errors; review it before using it.

Analyse a video

In Analysis, choose Ask about a video. Add an MP4 clip up to two minutes, 1080p and

100 MB. Use the preview to check it, then choose Summarize this video or ask a question

in English. Choose a response length and Generate.

The price hint is a conservative estimate. The final charge follows actual usage and may be

much lower. Saved answers can be copied or downloaded as text;

Tweak prompt restores the clip and model. Refresh results recovers an interrupted

request or delayed charge without starting another generation. Switching between video,

image and audio clears the incompatible attachment. Review the answer before using it.

Write text or analyse an image

Choose Text to write a scene, script, or creative brief. Choose Analysis with Qwen 3.8 to ask a question

about an image or extract its visible text. This workflow accepts one PNG, JPEG, or WebP image up to 10 MB.

Write your prompt, choose Short, Standard, or Detailed response length, and click Generate.

Use the model menu to choose a Text model without losing your prompt. Nemotron expects English prompts.

Text and Analysis prompts are limited to 4,000 UTF-8 bytes; some characters use more than one byte.

Qwen 3.8 image Analysis resizes its analysis copy to fit within 1,024 pixels; your original stays unchanged. The dedicated picture workflows described above retain the full source instead.

The price shown before generation is a maximum estimate. Your final charge follows actual token usage.

Saved results show their model below the composer. Very small charges retain enough decimals to show

the amount. Final charge pending means your output is ready but its bill has not arrived yet.

Refresh results immediately checks the existing request for its final charge. It does not

start another paid generation. You can still copy or download the output while billing is pending.

Copy text copies the result; Download .txt saves it.

Tweak prompt restores the recipe for review before another paid generation. Responses can stop at

the selected length; results reported as incomplete carry a notice.

Refresh results checks existing request IDs after an interrupted session. It does not start another

generation. If a request is still pending, check History before retrying. Text results remain separate

from media sent to projects or shared to the gallery.

On wide screens, the library shows ten columns; smaller screens use fewer readable columns.

Each page holds up to 100 generations in either view. Settings opens within the screen.

Measured video lengths are displayed to two decimal places.

Shape the result

For images, Format chooses the aspect ratio or custom dimensions, and How many chooses the number of images. These are separate controls beside the model selector. Other categories use Output settings for their essential controls. Use + to upload a reference. What the controls offer depends on your model.

• A reference. Some models accept a picture to work from: Add a reference image for an

  image, Add a first frame for a video, Build from a picture for a 3D object. Meshy 7.1

  offers Build from pictures: add a first photo, then Add another angle for up to four

  photos of the same object from different sides. Meshy 7.1 Multi-image requires two to four views;

  Generate stays unavailable until the second view is attached. With pictures attached, your description is optional and guides

  the surface rather than the shape. Removing the first picture promotes the next angle.

  Run again keeps the uploaded source views; if those views are unavailable, attach them in 3D

  before generating again.

  Sound made for a video needs Add the video.

• Surface. For a 3D mesh, choose Textured or Geometry only. Meshy 7.1 supports both

  choices whether you start from a description or pictures.

• Rig. Meshy 7.1 offers Plain object or Rigged character. Rigging is meant for

  human-shaped characters and returns GLB or FBX, ready to animate in another application.

  If a rig cannot be made, the plain mesh is returned. The estimate includes requested rigging;

  the final charge follows the work billed. Billed work may still be charged if a job fails or is

  canceled; check History for the final amount. Changing to an engine that accepts fewer pictures

  keeps the first and removes extra angles, with a notice.

• Model. Search the model menu or choose a favourite. Browse the complete catalog by category; models awaiting verification cannot generate yet.

• Shape, Duration and Resolution set the size and length of the result. Video Duration

  offers every supported whole second, or the model's fixed lengths. Grok Imagine Video 1.5 starts

  at 1 second; Seedance 2.5 starts at 4 seconds. Switching models keeps a supported selection;

  otherwise it moves to the nearest supported length and tells you.

• How many makes several images from one prompt. The estimate multiplies with the count.

• Lyrics for music, and Delivery for a voice line.

Kling 4.0 video

Video adds Kling 4.0 Text, Kling 4.0 Image and Kling 4.0 Reference, with 3 to 30 seconds at 720p or 1080p. Text starts from your description. Image needs a first picture and optionally an ending picture. Reference accepts a first frame, up to ten pictures including that frame, five videos, and seven character or object elements. Each element uses either a front picture and one to three extra pictures, or a video. Up to three elements can use video. Each video must be 3 to 30 seconds, and all reference and element videos together must fit 30 seconds. Use the prompt helper to insert @Image1, @Video1 or @Element1 for the reference you mean. Saved recipes restore every reference through Tweak. Older Kling models remain in the Sandbox.

Choose Kling 4.0 Edit under A video to edit to change an existing clip. It accepts 3 to 30 seconds, prepares small clips to a 720-pixel short side and renders at 1080p. The prepared clip must be at most 300 MB, 700 to 4,553 pixels per side, between 2:5 and 5:2 in shape, and run at 24 to 60 frames per second. Its estimate includes input plus output length, with a minimum of twice the output length. Video references and element videos also affect Reference estimates. Hover or focus Generate to check the Estimated cost; final charges follow actual usage.

There is no sound switch or keyframe control for Kling 4.0 in this interface.

Editing a video

Under Video, Start from switches between writing a prompt for a new video and changing an

existing clip.

• Set Start from to A video to edit, then click Video to edit and choose a clip — an

  MP4, MOV or WebM, of any length or size. MOV and WebM, and any clip too large to send as it

  is, are converted to MP4 in your browser before anything is uploaded; if your browser cannot do

  that conversion, export the clip as MP4 and try again.

• Under What should change?, describe the change you want, in place of the usual prompt.

• The clip sets the length and size, so Shape, Duration, Resolution and

  How many are hidden while you edit a video. Model stays available: choose FLUX 3 Video Edit

  to edit the attached clip.

• Click Generate. The result is a new video; your original clip is kept.

  FLUX renders 720p.

Longer clips

The clip can be any length. If it is 15 seconds or longer, a trimmer appears under it: drag the

highlighted part, or its two handles, to choose up to 14.5 seconds, use Play the part to check

it, then click Use this part. Only that part is converted, priced and sent — so a shorter part

costs less. Change the part takes you back to the trimmer.

FLUX accepts under 50 MB once converted. A part that is still larger is not sent: the

trimmer stays open and asks you to Choose a shorter part.

Generate's Estimated cost appears once your clip has finished preparing, and is priced by the

clip's length. The edit itself is charged for the length of the video it produces, which is

normally the same length as your clip.

Every video result also offers Edit this video, which opens this mode with that clip already

loaded — see Your sandbox, below.

Sound FX from a video

Under Sound FX, set Make sound for to A video, then click Add the video and choose

a clip — an MP4, MOV or WebM, any length. MOV and WebM are converted to MP4 in your browser

before anything is uploaded; if your browser cannot do that conversion, export the clip as MP4 and

try again.

If the clip is 60 seconds or longer, a trimmer appears under it: choose up to 59.5 seconds and

click Use this part. Only that part is converted, priced and sent, and Change the part

takes you back to the trimmer. What is sent must be under 100 MB once converted; a part that is

still larger asks you to Choose a shorter part.

• Duration is hidden while you score a clip — the clip sets the length, so there is nothing to

  choose.

• Describe the sound (optional) lets you steer the result — "footsteps on gravel, distant

  thunder". Leave it empty and the sound is made from the clip alone.

• Click Generate. The result matches what the clip shows.

Generate's Estimated cost appears once your clip has finished preparing, and is priced by the

clip's own length.

A result scored from a video offers no Run again, Tweak or Variations — it has no clip

of its own to run again. To score it again, open the video it was made from and use Add sound

there instead — see Your sandbox, below.

Eleven speech

Voice opens on Eleven v4, the quality-focused speech model. Choose Eleven v4 Turbo

from the Model menu for the faster variant. Both accept the existing delivery controls

and audio tags. ElevenLabs TTS v3 remains available from the full menu.

Voice library

The voice library is available to every signed-in user when the voice connection is ready. Custom voice creation is unavailable.

With an Eleven speech model selected, open Output settings → Voice library. Explore ready-made voices and filter by language, gender, age, accent or use case. Favorites use the same filters. Only voices matching the selected filters remain visible. The star at the top-right of a voice card adds or removes it from Favorites. An existing sample plays from the audio control below the description, without generating new audio. Open Description when you want more detail about a voice.

The selected saved voice reads your next line. Tweak and Run again preserve that identity, your writing and the requested seed. Switching models clears a saved-voice selection so an incompatible identity cannot carry over.

For supported speech, music and sound effects, Sandbox checks compatible connections and uses the lower verified generation cost. Confirm the current estimate before submission. The final charge follows actual usage and the exchange rate in that quote; a failed or interrupted request can still have a charge. Check History before retrying a request that needs recovery.

Gemini speech

Choose Voice, then Gemini 3.8 Flash TTS or Gemini 3.8 Flash Lite TTS from the

Model menu. Enter the words to speak and choose a Delivery. Both use the Kore voice

and return WAV audio. You can include <laugh> or <sigh> in the spoken text.

Delivery directions are separate from the spoken words and do not add to the billed text.

Check the Estimated cost before generating; open the result to play or download the audio.

Voice replies with Grok Voice

Choose Voice, then Grok Voice from the Model menu. Attach an audio recording up to

10 minutes and 50 MB. Instructions (optional) can guide the reply; leave them empty to use

the model's default. Choose a built-in Voice, or leave Default selected.

Click Generate to receive an audio reply. Open the result to play or download it and read

its transcript when available. Tweak and Run again reuse the uploaded recording and voice.

Web search and external tools are unavailable for these replies.

The Estimated cost uses previous replies. The final charge uses metered usage and may arrive

after the audio; check History for the final amount. A refused reply may still incur a charge

for billed usage.

If an estimate is temporarily unavailable, use its retry button to read the price again. This refresh does not start a generation or charge your balance. Generation stays disabled until a price is available.

Your sandbox

The results section follows the category selected in the composer: Image shows only images,

Voice shows only voice results, and each other category shows its own creations. Switch

category to see those saved results. Search reads prompts in the current category; changing

category clears the search and returns to the first page.

Use Sort by above Search for newest first, oldest first or model name order. Narrow the list

with Model, Status, From and Until. Date filters include both selected dates (UTC).

Filters search your complete saved history, including older models. Use Clear filters to reset them.

Pages contain up to 100 generations. Page numbers, Previous and Next appear above and below

the results, together with the displayed range. Choosing another page returns to the top of the results.

On smaller screens, fewer columns can mean more than ten rows. The composer stays at the bottom

of the window while results scroll. Account pages have no website footer.

The Grid view and List view buttons sit to the left of Search, on the same toolbar row.

The list shows each result's prompt, engine, shape, seed and cost.

Click a result to open it. The preview shows the prompt and what it was charged, and offers:

• Run again — the same settings, a new result. It starts a new paid job straight away. An

  edited video offers Edit this video instead, and a result scored from a video offers

  Add sound instead, since neither recipe has a clip to resubmit.

• Tweak — the same settings, loaded into the composer for you to change. Also absent from an

  edited video or a result scored from a video, for the same reason.

• Variations — four more images close to this one. It starts straight away, and is paid.

• Animate this and Make it 3D on an image, Add sound and Edit this video on a

  video. These fill the composer with the result as the reference; nothing is charged until you

  click Generate.

• Download — save the original file to your computer. Image, video, audio and 3D downloads keep

  your results open; downloading does not generate again.

• Remove — delete the result from your sandbox. There is no undo.

Use the left and right arrow keys to move between results in the selected category, and Esc

to close the preview.

Send a result to a project

1. Open the result and click Send to project, then choose the project.

It arrives in that project's assets, ready to place on a timeline. You can also go the other way —

from inside the editor, see Bringing media in.

Share with the community

Share a completed image or video from its preview or library actions. The sharing dialog lets you add an optional title and up to eight comma-separated tags before choosing Submit for review. The asset and its details appear publicly only after approval. See

Community gallery.

Virtual try-on and product holding

Open Templates in your account menu, then choose Virtual try-on or Product holding.

These make an image without opening a project.

Choose a Person image, then add one to three references. Virtual try-on accepts garments and

accessories; Product holding accepts product photos, packaging, other angles or companion items.

Each image can come from an upload, a publicly accessible image URL, your Sandbox results or a

project's generated images. To use a link, paste it into the image URL field and click Use image URL.

Instructions are optional. Describe which clothes to replace or keep, how to layer them, or

which hand should hold a product and which way its logo should face. Under Advanced settings,

choose an aspect ratio or leave Keep source ratio selected. The Seed starts at 5555; reuse

it with the same images and settings to repeat a result.

Hover or focus Generate to see its estimated cost, then click it to make one image. Inputs

stay locked while it runs. Cancel stops a queued or running job; cancellation is unavailable

while the completed image is being saved. Your result replaces the example and is recorded in

Your sandbox. Download image opens the output for saving.

Use these settings copies the example's instructions, aspect ratio and seed. It keeps your

selected images, so it does not reproduce the example unless you supply the same sources. Template

results cannot be replayed in the general Sandbox composer; open their template to make another.

Seedance 2.0 US 4K is available in the Sandbox model list with fixed 4K output. In Projects it appears only when the project size is 4K (3840 × 2160); changing to a smaller size switches its editable selections to Seedance 2.5. Generation requires verified 4K pricing to be available.

Results from the Clean audio template

are saved here too. Open in Templates restores the recording and settings; checking an

interrupted cleanup reads the original job instead of generating again.

More voice languages

Choose Kokoro Hindi, French, Spanish, Italian, Brazilian Portuguese, Japanese or Mandarin Chinese in Voice to speak text in that language. The new language presets are Sara (Italian), Dora (Brazilian Portuguese), Alpha (Japanese) and Xiaobei (Mandarin Chinese). Each model uses its matching preset. The estimate reserves whole character blocks; the final charge follows supplier billing and may remain pending briefly while the audio is already ready.

Hunyuan Rapid textured workflows stay hidden while their outputs are being verified. The model menu and Favourites show available workflows only.

OBJ results

OBJ results show a geometry preview with a download for the mesh. Supplied material and texture

files have separate download links. The preview does not show materials; the supplied files may

omit some texture maps. A missing preview does not remove the downloads.

Instrumental music cues

In Music, choose Stable Audio Small Music Base or Stable Audio Medium Music Base.

Describe your instrumental cue, then open Output settings to choose its duration. Small

offers lengths up to two minutes; Medium up to six minutes twenty seconds. These workflows

return WAV audio and have no separate lyrics input. Switching to a model with a shorter

limit adjusts the duration and shows a notice.

Play or download the saved audio, or use Tweak to restore its model, description and duration.

The final charge follows the supplier bill. If it is pending, Refresh results checks the

existing job without generating another cue.

Background removal results

1. The Library retains original AI cutouts and separately saved finished variations from Remove background. Open in Templates restores the original source, settings and transparency mask for further editing, including variations with a solid background. Check original job retrieves an interrupted AI request without another paid submission. Local color removal and saved refinements do not create a billed generation job.

Relight results

Relight videos made in Templates or a project are saved in the library with their source, lighting reference, settings and returned seed. Open in Templates restores the recipe. If a submitted job loses its connection, check that existing job before explicitly starting another paid generation.

Hunyuan Video 1.5

Choose Hunyuan Video 1.5 Text for a description, or Hunyuan Video 1.5 Image for a description and one first-frame picture. Both make silent 480p video at the model’s default length, in landscape or portrait. There are no last-frame, audio or custom-length controls. Tweak restores the model, description, format and source picture. The final charge follows the verified output bill; refreshing a saved result checks its existing job.

Standard square images

Janus makes 384 × 384 square images; Format shows those exact pixels. More verified 1K choices include FLUX General, FLUX SRPO, Lumina Image V2, Recraft 20B, Sana 1.5, Pony V7, CogView4, OmniGen, Ovis Image, DreamShaper, Fast LCM, BITDance, Juggernaut FLUX Base and RunDiffusion Photo. LoRA-capable choices use verified base or built-in presets, including FLUX Kontext, Ideogram V4, Z Image Base and ERNIE Image. Custom adapters and guide-image workflows are available only when their required controls have been verified. Recraft 20B descriptions are limited to 1,000 characters. Processing-based models show estimates; History records the final charge from actual billed work.

Additional 1K square presets include FLUX Dev LoRA, FLUX 2 LoRA and Realism, Krea 2 Turbo LoRA, Qwen Image 2512 LoRA, Z Image Turbo LoRA and Tiling, FLUX 1 Dev and Krea, FLUX Klein LoRA variants, Cosmos 3 Super Image, Juggernaut FLUX Lightning and Pro, Playground 2.5, Realistic Vision, Boogu Image, Stable Diffusion 1.5, GLM Image and HiDream O1 Image Dev. These use their verified base or built-in style presets; custom adapter files and reference controls are not exposed by these workflows. Compute-priced presets show estimates and settle from actual supplier bills. Wan 2.5 Image produces fixed 1440 × 1440 square images; Format shows its exact pixel size. After a failed or interrupted request, check History for the final charge. A failed result alone does not confirm zero billing. Wan 2.5 descriptions are limited to 2,000 characters. If a description exceeds its model's limit, Generate stays unavailable and shows the limit; your text is preserved so you can shorten it or choose another model.

Recraft V4 Pro, V4.1 Pro and V4.1 Utility Pro, plus Seedream 5 Flash and Lite, offer 2K square images. Recraft V4, V4.1, V4.1 Flash and V4.1 Utility offer 1K images. Other 1K choices include GPT Image 2 API, GPT Image 2.5 Flare, Ideogram V4 Fast and Instant, Sana Sprint, Stable Diffusion 3.5 Medium, LongCat Image, Hunyuan Image 2.1 and 3 Instruct, FLUX 2 Klein 4B and 9B Base, and Wan 2.7 Image and Image Pro.

Existing GPT Image 2 project choices keep using GPT Image 2.5 Sunburst. Choose GPT Image 2 API separately for the newly verified square-image workflow.

GPT Image 2 API and Flare show variable usage estimates; the final charge follows the actual supplier bill. Recraft, GPT and Seedream 5 workflows do not support a reusable seed. Their saved recipes restore the description, model and format without inventing a seed. A failed provider result does not become an available model. Billing refresh checks the existing job and never starts another generation.

Seedream 4 and 4.5 offer 2K square output. Ideogram V4.5, Qwen Image and Qwen Image 2512, FLUX 2 Flex, FLUX 1 Krea Dev, Hunyuan Image 3, Stable Diffusion 3.5 Large and 3 Medium, ERNIE Image Turbo and HiDream O1 Image offer 1K square output. Open Format to see the fixed size. How many chooses separate variations. Saved recipes preserve a returned seed; when a model supplies no seed, Tweak does not invent one. FLUX 2 Flex shows a metered estimate and finalizes against actual billed processing.

FLUX 2 Max, Flash, Turbo, Klein 4B and Klein 9B, Krea 2 Turbo and Ideogram V3 join SDXL, SDXL Lightning, HiDream I1 Fast, Dev and Full, Kolors, AuraFlow and FLUX.1 Schnell. These models offer a standard square output from a description. These workflows have no source-picture or custom-size controls. Use How many for separately generated variations. Tweak restores the model and description; the final charge follows the completed job. Metered workflows, including FLUX 2 Max and compute-priced models, show an estimate before generation; their final charge can differ.

The existing FLUX 2 entry also offers other formats and picture editing. Description-only generation uses Pro; adding pictures switches to Dev editing. Add one picture, then use Add another picture for up to four references. Removing the first picture keeps the others. Tweak restores all saved references in their original order. Switching to a model that accepts one picture removes the extra references and tells you why.

The estimate includes every reference and output for each variation. Adding references increases the estimate; two variations include the input work twice. Hover or focus Generate to check the current estimated cost and service fee. Final charges follow actual supplier usage.

Make an image in a reference style

Choose Recraft V4 Style, Recraft V4 Pro Style, or either Style SVG version in Image. Add one PNG, JPG or WEBP picture with Add a style reference, then describe the picture you want to make. Generate becomes available once the required picture and description are ready.

These workflows use a square canvas. Raster models produce 1K or 2K WEBP; SVG models produce scalable vectors. How many controls separate generations. Each estimate includes the supplier's style creation fee for the attached reference.

Tweak restores the original reference, description, model and settings. Download preserves the original WEBP or SVG. Video, 3D and project actions prepare a PNG copy; the original SVG remains in your results.

SVG vector images

In Image, choose Recraft V4 SVG, Recraft V4.1 SVG, or either Pro SVG version. Describe a picture and choose Generate. These models produce a scalable vector on a square canvas; Format shows SVG · 1:1, while How many remains a separate choice.

Open a saved result to preview it. Download saves the original .svg without converting it to PNG or leaving your results. Tweak restores the model and description; these models do not return a reproducible seed. Check the estimate before generating variations.

Choose Animate this, Make it 3D or Send to project to use a PNG copy of the saved vector. Preparing image appears during conversion and upload. Video and 3D open their composer with that image attached; choose a model and check its estimate before pressing Generate. The conversion itself starts no generation. Sending to a project saves the PNG in its assets. Download always keeps the original SVG. If preparation fails, the result stays open so you can retry.

Community gallery sharing remains unavailable for SVG results.

Edit an image or follow its outlines

FLUX 2 Pro Edit, Max Edit, Flex Edit, Flash Edit, Turbo Edit, the Klein 4B and Klein 9B editing variants, and FLUX 2 LoRA Edit also take one source picture and a description of the change. These presets upload your full picture without cropping and produce one 1024-pixel square per generation. The model may adjust input resolution. Choose a PNG, JPG or WEBP file up to 32 MB and 8192 pixels on each side. How many makes separate variations. The LoRA variants use their standard preset; there is no custom adapter control.

Their saved recipes retain the uploaded source picture. Tweak restores it, and Download keeps the original generated PNG. The price hint is an estimate; the final charge follows actual supplier usage.

FLUX Dev Image to Image uses source dimensions up to 2048 pixels on each side; a larger source can change the output shape. FLUX Kontext Dev preserves the source shape up to 1536 pixels on its longest side. They do not show a shape or resolution control that the model would ignore. Estimates account for source dimensions and supported processing sizes. Final charges follow the actual supplier bill. Add a source picture before checking the estimate for these models and Pro, Max or Flex Edit.

Choose Qwen Image Edit, Qwen Image Edit Plus, Qwen Image Edit 2511, Z Image Turbo Edit, FLUX Kontext Pro or FLUX Kontext Max in Image. Use Add a source picture for one PNG, JPG or WEBP file, then describe the change you want. Generate stays unavailable until the picture and description are ready.

For SDXL ControlNet Union, the source picture’s outlines guide the new image. Describe what should fill those outlines. These workflows use a fixed square canvas; Format shows its size and How many chooses separately generated variations. Custom adapters and additional guide types are not part of these presets.

Tweak, Run again and Variations restore the source picture. Saved recipes keep a returned seed without inventing one. Download preserves the original PNG or JPEG. Quotes are estimates; the final charge follows the supplier’s actual bill, including compute-priced ControlNet processing.

Seamless materials

In Image, choose PATINA Material to describe a surface, or PATINA Material Extract to add a PNG, JPG or WEBP picture and describe which surface to extract. These workflows make a square 1024 × 1024 seamless texture. How many chooses separate generations.

Open the result to find Material files: the texture, base colour, normal, roughness, metalness and height maps. Each button downloads that original PNG with a filename identifying its map, while keeping the result open. Use these files in your 3D software. Tweak restores the description and, for extraction, the source picture. Video, 3D and project actions use the main texture image.

The estimate covers the complete workflow. Final charges follow actual supplier usage and can differ from the estimate. After a submitted failure, check History for the final amount before generating again.

FLUX Krea LoRA Stream

Choose FLUX Krea LoRA Stream in Image for a square 1024 by 1024 image. The finished result is an original JPEG; download it from its preview without leaving the Sandbox. This preset uses the base model, with no custom adapters.

Saved results retain the final image and its returned seed. Tweak restores the description and controls. History uses the same verified final image. The estimate is shown before generation, and the final charge follows actual supplier usage.

Material maps from a picture

Choose PATINA Image Maps in Image and add one PNG, JPG or WEBP picture. No description is needed. The whole picture is fitted to at most 1024 pixels on its longest side, without cropping or adding a border; smaller pictures are not enlarged. Choose a picture up to 8192 pixels on each side and 32 MB.

The model chooses dimensions from the prepared source, so Format has no shape or custom-size choices. How many creates separate generations. Open a result to download its five original PNG maps: Base colour, Normal, Roughness, Metalness and Height. There is no additional texture file in this workflow.

Tweak, Run again and Variations restore the prepared source picture and returned seed. The estimate covers the complete bounded workflow; the final charge follows actual supplier usage. Check History if billing is still pending.

Remove a picture background

Choose Image, then BRIA Background Removal, Ideogram Remove Background, Pixelcut Background Removal, BiRefNet, or BiRefNet V2. Add one picture; these workflows do not need a description. They preserve the source dimensions, up to 8,192 pixels on either edge and a 32 MB prepared PNG. Generate makes one cutout at a time.

Open the result to see transparency against a checkerboard. Download saves the original transparent PNG. Tweak restores the source so you can replace the picture or choose another model. Sending the result to a project carries its transparency with it.

BiRefNet prices are estimates: the final charge follows the supplier's actual compute usage, which can vary even for pictures of the same size. Check the estimate before generating and use History to inspect the final charge.

Guided background removal

Choose BRIA Guided Background Removal in Image and add a picture. Use What should stay? (optional) to keep specific subjects, such as “the person with the mat and the plant” or “only the chair.” Leave it empty for the obvious subject.

The result keeps the full source canvas and original pixels with a new alpha channel. Download saves the transparent PNG. Tweak restores the source and instructions. Each request makes one cutout; this model has no seed or output-size controls.

Selectable voice presets

Choose Gemini 3.1 Flash TTS or MiniMax Speech 2.8 Turbo in Voice. Open settings to select a Voice preset and delivery. Gemini offers thirty native presets; MiniMax offers eight. Write the words to speak, then Generate. The composer shows the selected model’s text limit and keeps overlong text so you can shorten it.

Gemini downloads the original WAV; MiniMax downloads the original MP3. Tweak restores the saved voice and delivery. Estimates count spoken prompt characters, including inline expression markers. Separate delivery instructions are not billed as spoken characters. Final charges follow actual supplier billing and appear in History after settlement.

More MiniMax speech choices

Voice includes MiniMax Speech 02 HD, 02 Turbo, 2.5 HD, 2.5 Turbo, 2.6 HD and 2.6 Turbo. Choose a Voice preset and delivery in settings, then write the words to speak. The 02 and 2.5 models accept up to 5,000 characters; 2.6 accepts up to 10,000. Overlong writing stays in the textbox so you can shorten it.

These models download their original MP3 and save the chosen voice and delivery for Tweak. Estimates use spoken characters at the selected model's tariff; final charges follow the supplier bill in History.

More preset speech voices

Choose ElevenLabs Multilingual v2, ElevenLabs Turbo v2.5, xAI Text to Speech or Qwen 3 Preset Speech in Voice. Choose a Voice preset, then write the words to speak. ElevenLabs offers Rachel; xAI offers its native voice list, and Qwen offers nine presets. Qwen also offers delivery choices. These workflows keep the original MP3, and Tweak restores speech and the chosen preset.

ElevenLabs and Qwen accept up to 5,000 characters; xAI accepts up to 15,000. Qwen pricing rounds speech and delivery instructions together up to the next 1,000 characters. For example, 996 spoken characters plus the five-character happy instruction use two blocks. Other models estimate fractional character usage. Final charges follow the supplier bill in History.

Voice result previews show Format. Preset workflows without a seed omit that control.

More native speech choices

Qwen 3 Preset Speech 0.6B offers nine presets and delivery choices, with up to 5,000 characters of speech. It keeps the original MP3. Like the larger Qwen preset model, its estimate rounds speech and delivery instructions together up to the next 1,000 characters.

Async Speech Pro offers 100 native voice presets and accepts up to 5,000 characters. It keeps the original MP3 and estimates whole 1,000-character speech blocks.

Qwen Audio 3 Flash Speech offers 45 presets and accepts up to 2,000 characters. It keeps the original WAV and estimates fractional character usage. Tweak restores saved writing and voice. Character limits count complete Unicode characters, including emoji; overlong writing stays in the textbox so you can shorten it. Final charges follow the actual supplier bill.

Verified native preset speech workflows allow up to fifteen minutes for queueing and generation. If a request is canceled, History can show pending billing until the supplier receipt arrives.

Native music and lyrics

Choose MiniMax Music 2.5 or MiniMax Music 2.6 in Music. Describe the style, then optionally add lyrics in settings. Leave lyrics empty for an instrumental. The model chooses the track length, so these models have no duration control. Lyrics are submitted as written, without automatic rewriting. The composer keeps overlong writing so you can shorten it.

ElevenLabs Music v2 makes instrumental music from a description, with selectable duration. Estimates include the original MP3’s extra audio frame before rounding to whole minutes. A request at an exact minute boundary can therefore be billed for another minute; check the estimate before Generate. Final charges always follow actual supplier billing.

Download preserves the original MP3. Tweak restores the description, supported duration and lyrics. Switching to instrumental-only music keeps unused lyrics in the composer’s memory; they are excluded from that request and its saved recipe. These workflows do not expose a seed.

Native Lyria music

Choose Lyria 2 in Music and describe the instrumental track. Its output is about thirty seconds; the model determines the exact length, so there is no duration or lyrics control. Download preserves the original WAV. Tweak restores the description and saved recipe.

ElevenLabs Music v2.5

Choose ElevenLabs Music v2.5 in Music for an instrumental track from your description. In Output settings, choose a duration from three seconds to ten minutes. This workflow has no lyrics or seed control.

The original MP3 includes an extra audio frame. Estimates include that frame before rounding to whole minutes, so an exact-minute request can cost another minute. Check the estimate before Generate; final charges follow actual supplier billing.

Play or download the original audio after returning, or use Tweak to restore the description, model and requested duration.

Mureka music and lyrics

Choose Mureka 9.5 Instrumental in Music for a track without vocals, or Mureka 9.5 Song for a song. Both preserve the original MP3 and let the model choose the length; they have no duration or seed control.

For a song, add your lyrics in settings or describe a song and let the model write them. Supplied lyrics can be used without a description. With lyrics, the description provides musical style. The two modes have different estimates: check the price after changing the lyrics. Estimates follow published supplier prices, while the final charge follows the actual supplier bill and may be lower.

Choose Mureka Lyrics in Text to generate a title and lyrics without audio. Its price is per request, so there is no response-length control. Saved results remain available after returning; Tweak prompt restores the original description and model.

DiffRhythm songs with timed lyrics

Choose DiffRhythm in Music and open Output settings. Choose 95 seconds or 285 seconds, then add timestamped lyrics. A musical style description is optional.

Start every lyric line with its time, such as [00:10.00]Your first lyric line followed by [00:16.00]Your next lyric line. Add at least two lines, use increasing times and keep every time before the selected duration. Lyrics are submitted as written. If a time needs fixing, the composer keeps your writing and explains the correction before generation.

The original WAV may run slightly longer than the selected duration. Play or download it after returning, or use Tweak to restore the style, timed lyrics and duration. Check the estimate before Generate; final charges follow actual supplier billing.

Stable Audio Open music

Choose Stable Audio Open in Music and describe an instrumental cue. In Output settings, choose a duration up to 47 seconds. This workflow has no lyrics or seed control.

The estimate is based on the supplier's average cost per generation. Final charges follow the actual compute time used, which can differ from the length of your audio. Generate stays unavailable until an estimate is ready.

Play or download the original WAV after returning. Tweak restores the description, model and selected duration.

Music variations from an audio source

Choose Stable Audio Small Music Variation, Small Music Base Variation, Medium Music Variation or Medium Music Base Variation in Music. Add an MP3 or WAV through +, listen to the source preview, and describe the music you want to create.

Source music can be up to 380 seconds and 100 MB. In Output settings, choose a length up to 120 seconds for Small models or 380 seconds for Medium models. These workflows make one original WAV and have no lyrics control.

The uploaded source is saved before generation can start. Tweak restores that source, description and duration. Play or download the original result when you return. Final charges follow the supplier's actual bill per audio output.

Edit or extend source music

Choose a Stable Audio Music Region Edit model to replace part of an existing MP3 or WAV. Add the source through +, listen to its preview, and describe what should play in the selected region. In Output settings, set From and To in seconds. The end must be after the start and within the source. Download keeps the full original WAV result.

Choose a Stable Audio Music Extension model to add music before or after a source. In Output settings, set At start and At end in seconds; at least one must be greater than zero. The source and added time together must fit within 120 seconds for Small models or 380 seconds for Medium models.

These workflows make one audio result, with no duration or lyrics control. The source must be MP3 or WAV, no larger than 100 MB, and fit the selected model's full-track limit. Source playback must load before Generate becomes available. Tweak restores the source, description and editing settings. Check the displayed estimate before generating; History shows the final charge once usage billing completes.

Songs from written lyrics and reference music

Choose MiniMax Music 2, describe the style, and add your required Lyrics in Output settings. Both inputs need at least ten characters. The description can be up to 2,000 characters and lyrics up to 3,000. The model chooses the song length and produces an original MP3.

Choose MiniMax Music 1.5 to write Lyrics in the main textbox and a separate required Music style in Output settings. Lyrics need 10 to 600 characters; style needs 10 to 3,000. The original MP3 uses a model-chosen length.

Choose MiniMax Reference Song to write your lyrics in the main textbox and add an MP3 or WAV song through +. The source must contain music and vocals, be longer than 15 seconds, and fit the Sandbox's 380-second and 100 MB source limits. The source preview must load before Generate becomes available. The uploaded source is saved before generation starts; the model chooses the result length.

ACE Step Music accepts a style description, optional lyrics and a duration from 5 to 240 seconds, producing an original WAV. Sonilo 1.1 makes one original AAC/M4A track from a description, with duration choices up to 600 seconds. Tracks shorter than ten seconds are billed for ten seconds; the estimate includes that minimum before generating.

Play or download the original audio after returning. Tweak restores the original writing, supported controls and any reference song. Music previews show the native file format and omit unsupported seed information. History shows the final charge once actual usage billing completes.

Seed Speech voices

Choose Seed Speech 2 in Voice and open Output settings to select a voice preset. Voice names show their supported languages. Write up to 5,000 characters; overlong writing stays in the box so you can shorten it.

Generate returns MP3 speech. Download keeps the original file, and Tweak restores your words and selected voice. This model has no seed or delivery controls. Estimates use the exact spoken character count; History shows the final supplier-based charge.

Short narration with Chatterbox HD

Choose Chatterbox HD Short Speech in Voice for writing up to 500 characters. Open Output settings to choose one of nine voice presets. Use another Voice model for longer narration.

Listen to the saved WAV, download the original, or choose Tweak to restore the text, voice and seed. Overlong writing stays in the box so you can shorten it. Estimates include the supplier’s minimum character block; final charges follow the actual bill.

English narration with VibeVoice

Choose VibeVoice 0.5B Speech in Voice for English narration up to 5,000 characters. Open Output settings to choose Frank, Wayne, Carter, Emma, Grace or Mike.

Play the saved MP3, download the original, or choose Tweak to restore your words, voice and known seed. Overlong writing stays in the box. Estimates use script length; the final charge follows the supplier’s billed audio and appears in History.

Short two-speaker dialogue

Choose Dia Short Dialogue in Voice for an English conversation of up to 250 characters, including speaker markers. In Output settings, use Speaker 1 and Speaker 2 to append [S1] and [S2] to your writing. Write each speaker’s line after its marker.

The model chooses both voices; there are no voice presets or seed controls. Keep the conversation short. Overlong writing stays in the textbox so you can shorten it. Download keeps the original WAV, and Tweak restores the exact dialogue. Estimates include one full 1,000-character block; History shows the final supplier-based charge.

More square image models

Choose Nano Banana 2, Nano Banana 2.1, FLUX 3 Image, GPT Image 1.5 or Gemini 2.5 Flash Image in Image. Describe your scene and generate a square 1024×1024 PNG. These workflows use verified output presets and accept no source picture.

Descriptions can contain up to 2,500 characters. Nano Banana and Gemini need at least three characters; GPT Image 1.5 needs two. FLUX 3 Image and GPT Image 1.5 have no seed control. GPT Image 1.5 uses its High quality preset; Nano Banana 2 uses 1K without web search or thinking, while Nano Banana 2.1 uses minimal thinking.

Token-priced models show a cost estimate. The final charge follows the supplier’s actual bill and your account fee. Download preserves the original PNG, and Tweak restores your writing and supported settings.

BRIA, Wan and transparent image generation

Choose Ideogram V3 Transparent to create an isolated object from a description without uploading a picture. Its original PNG retains transparency; the checkerboard in the result preview helps you see the transparent area. Download keeps the original file, and Tweak restores your saved description and model.

BRIA Fibo, Fibo Lite, BRIA Image Base, Fast, HD and Wan 2.2 A14B Image also generate from a description, without source pictures. These presets support seed. Each generation returns one square image: BRIA HD uses native 1536 × 1536 JPEG, while the other presets use 1024 × 1024 PNG. Choose How many separately when you want several generations, and check the estimate before generating. History shows the final supplier charge.

MAI, Luma Uni and more image engines

MAI Image 2.5, MAI Image 2.5 Pro, Phota, Nucleus Image and BRIA Fibo Gen 1.5 make square 1024 × 1024 images. Luma Uni 1 and Luma Uni 1 Max make square 2048 × 2048 images. Keep the original PNG downloads.

These description-only presets allow up to 2,500 characters. MAI and Phota need at least three characters. Search stays off, Luma uses its automatic style, Phota uses its NB2 preset and Fibo uses its native 1MP preset. Only Nucleus and Fibo expose a seed. MAI quotes remain estimates because text and image tokens are metered. History shows final supplier-based charges, and Tweak restores the original writing and supported settings.

More native square image models

Image adds Gemini 3.1 Flash Image, Grok Imagine Quality, GPT Image 1 High, Kling Image 3 Text, Kling Image O3 Text and MiniMax Image 01. Each makes one square 1024 × 1024 image from your description. MiniMax keeps its original JPEG; the others keep original PNG. Only Gemini 3.1 Flash Image supports a seed.

MiniMax allows up to 1,500 characters; the others allow 2,500. Gemini needs at least three characters, and GPT needs two. Search and prompt expansion stay off; Gemini Flash uses minimal thinking and GPT uses High quality. GPT estimates include the native High-quality minimum and the supplier’s whole-cent rounding. History shows final supplier charges. Use Tweak to restore the exact writing and supported settings.

Luma Photon image presets

Luma Photon and Luma Photon Flash make one square 1536 × 1536 image from your description. Keep the original JPEG, or restore the exact saved writing with Tweak. These models have no seed or source-picture controls. Write between three and 2,500 characters. The fixed preset estimate covers the native image; final charges follow the supplier bill in History.

Ideogram variants and Kontext Max

Image adds Ideogram V2a Turbo, Ideogram V2 Turbo, Ideogram V2a and FLUX Kontext Max Text. Each makes one square 1K image from your description, with original PNG downloads. Prompt expansion stays off. Use Tweak to restore the exact saved writing and seed; History shows the final supplier-based charge.

More original and compact image models

Image adds Nano Banana Original, Nano Banana Lite, Grok Imagine Original, GPT Image 1 Mini, Ideogram V2, Krea 2 Medium and FLUX Kontext Pro Text with square 1024x1024 output. FLUX Pro 1.1 Ultra uses square 2048x2048 output. All eight verified presets return original PNG images from a description, without a source picture.

Use up to 2,500 characters. Nano Banana needs at least three characters; GPT Mini needs two. Emoji count as individual characters for the minimum. Grok and GPT Mini have no seed control. GPT Mini uses High quality, Krea uses its Medium creativity preset, and prompt expansion is disabled in Ideogram and FLUX presets.

Token-priced GPT estimates include the documented minimum for the selected High quality preset. The final charge follows the supplier's actual bill and your account fee. Download keeps the original image and Tweak restores the saved description and supported settings.

Native video presets

LTX 2.3 22B Text, Distilled Text and Quality Text generate landscape MP4 with audio at 1024 × 576, about five seconds at 24 fps. Hailuo 02 and 2.3 Standard Text generate silent landscape video at 1366 × 768; Pro Text uses 1920 × 1080. Hailuo chooses the native length, approximately six seconds. These description-only presets use model-default output settings without source pictures or a seed control.

LTX descriptions allow 2,500 characters and Hailuo allows 2,000. An overlong description stays in the textbox while Generate shows the limit. Open a saved result to play the original, check its actual dimensions, download its MP4 or restore the writing and model with Tweak. Final charges in History follow the supplier bill, including its billed duration.

Animate a first-frame picture

LTX 2.3 22B Image, Distilled Image and Quality Image, and Hailuo 02 and 2.3 Standard and Pro Image start from one picture. Choose a model, add your first-frame picture and describe the movement. Generate stays unavailable until both the picture and description are present. LTX descriptions allow 2,500 characters; Hailuo allows 2,000. Overlong writing stays in the textbox while the limit is shown.

Size follows the picture and the engine chooses native output. These presets have no shape, resolution, duration or seed choices. LTX produces MP4 with generated audio at about five seconds; Hailuo produces silent MP4 at about six seconds. The preview shows the actual saved dimensions. Download keeps the original video; Tweak restores the exact writing, model and uploaded first frame. In list view, these results say From picture rather than showing an unused format. Estimates are per video; final charges follow the actual supplier bill, including its billed duration.

Tripo H3.1 and Hunyuan3D V3

Tripo H3.1 Text and Hunyuan3D V3 Text make a textured GLB from a description of up to 1,024 characters. Their Image presets instead require one clear picture. Tripo H3.1 Multi-View requires two to four pictures of the same object: add the Front picture first, then Left, Back and Right views in order. A description is not needed for the picture presets.

These simple presets always return textured GLB. There are no seed, rigging, geometry-quality, surface or file-format choices. Switching to a description-only model removes pictures it cannot use. Your writing stays in the textbox if it exceeds the limit.

Open a result to inspect the original mesh. Its slow rotation follows elapsed time even on slower devices, and respects your reduced-motion preference. Download the 3D file keeps the original GLB. Tweak restores the exact model, writing and uploaded pictures; History shows the final supplier-based charge.

More Rodin, Hitem3D and Tripo presets

Rodin 2.5 Text generates a textured GLB from a description of up to 1,024 characters. Rodin 2.5 Image and Fast Image start from one object picture, with up to five views when you have them. The standard models use PBR materials; Fast uses baked shading. HighPack and other extra-cost options stay off.

Hitem3D 3.0 Image and Multi-View use the native 2048quality preset with texture and PBR. Tripo 2.5 Image and Multi-View use standard texture without PBR, style or quad options. Multi-View requires two to four pictures in Front, Left, Back, Right order. The picture presets need no description.

These simple presets always return textured GLB. Check the estimate before generating; final charges in History follow the supplier bill. Open the original in the rotating viewer, keep its original GLB download or use Tweak to restore the model and uploaded views. Each model’s view order is preserved, including Hunyuan’s existing Front/Back/Left workflow.

Automatic recovery and temporary results

When automatic recovery is available, cutipAI chooses the lowest verified estimate for your settings and can try another connection after a technical failure. The final charge includes provider-billed usage for each attempt plus the service fee.

Content refusals do not trigger another attempt. If submission is being checked, use History to recover the job instead of submitting it again. Automatic recovery is unavailable when there is no equivalent alternative for your settings.

If storage is full, generation can continue. Download a temporary result promptly or manage your allowance in Storage.

FLUX 3 video resolutions

1. FLUX 3 offers 720p, 1080p, 2K and 4K in Sandbox Video. The higher resolutions are finished with the video upsampler. These sizes work with text prompts, an opening image, first and last frames, and timed middle frames. Durations remain 5–20 whole seconds, with optional generated audio. The selected resolution changes the estimate and is retained with saved settings.

For changes to an existing clip, use Edit video. To enlarge an existing clip, use Upscale video.