k-anvas Get k-anvas

k-anvas

paint with a prompt.

A photo editor where a sentence works like a brush. Select any part of a picture, say what belongs there, and k-anvas fills it, removes it, extends it or lifts it out, on a GPU you run yourself.

Try the brush First release: coming soon

Desktop · iPad · iPhone

The k-anvas editor: a brass desk lamp, its base selected with a marquee, a new walnut base filled in, and the variant bar showing variant 2 of 3.
The k-anvas editor on desktop.

Try it here

Select it. Say it.

Brush over the lamp and press Remove, or switch to Fill and give it a new base. Your selection decides which pixels may change. Nothing outside it ever does.

Paint over the lamp

1/3
Operation

Remove what you select

A playback of real Qwen-Image-2.1 results from k-anvas's renderer tests, composited in your browser the way the app does it. In the app the same edit renders on your GPU in about 7 seconds.

Workflows

One composer. Every verb.

Generate, Fill, Remove, Expand, Extract and Compose all start the same way: a selection, a sentence, a button. Every result lands as a layer that remembers its recipe.

  1. Generate

    Start from a sentence.

    Describe a picture and get it at the size you need, up to about four megapixels. Words you ask for come out spelled right.

    4 to 5 s on the GPU at 1 MP

  2. Fill

    Change one part.

    Marquee the base, ask for a square walnut one, and get variants to choose from. The result is placed back through your selection, so the lamp, the desk and the light stay exactly as they were.

    6.6 s on the GPU

  3. Compare

    Choose with your eyes.

    See every variant side by side, hold to compare with the original, and keep the ones you like as layers. The rest wait in the Library, so nothing is lost.

    Variants · Compare · Keep as layers

  4. Remove

    Make it gone.

    Select Subject finds the lamp in one tap. Remove fills in the wall, the window and the desk that were behind it.

    6.6 s on the GPU

  5. Expand

    Give it room.

    Widen a square to 3:2 and the room carries on: more window, more wall, more desk. Your original pixels go back on top, untouched.

    10.4 s on the GPU at 1.5 MP

  6. Extract

    Lift it out.

    Extract returns the subject with real transparency, soft where the edges are soft. Drop it onto anything.

    6.6 s on the GPU, with alpha

  7. Compose

    Bring things together.

    Use up to ten pictures as references and say how they meet. k-anvas keeps track of which is image 1 and which is image 2.

    9.7 s on the GPU with two references

Straight off the renderer

Each of these took between 4 and 5 seconds of GPU time on the renderer k-anvas uses. The prompts are underneath.

  • A Swiss-style typographic poster titled FORMA, with a cobalt blue circle.

    “Swiss-style typographic poster … the title ‘FORMA’ in heavy grotesque letters” 4.0 s

  • An alpine lake at dawn with mist, snowy granite peaks and a rowboat.

    “Alpine lake at dawn, low mist over still water” 4.8 s

  • A rainy Tokyo alley at night with neon signs and a cyclist with a clear umbrella.

    “Rainy narrow alley in Tokyo at night … signs reading ‘喫茶 ミナト’” 4.5 s

  • A die-cut sticker of a cartoon fox reading a book, on a transparent background.

    “A die-cut sticker of a cartoon fox reading a book”, with transparency 4.0 s

  • A mid-century living room with concrete walls, an oak floor and a charcoal sofa.

    “Mid-century living room with board-formed concrete walls” 4.4 s

  • A wine glass half filled with red wine, on a transparent background.

    “A clear wine glass half filled with red wine”, with transparency 4.1 s

  • An overhead flat lay of white sneakers, a film camera, a notebook, earbuds and a watch.

    “Overhead flat lay on light grey paper” 4.6 s

  • A glossy cobalt blue cup on a saucer, on a transparent background.

    “A glossy cobalt blue ceramic coffee cup”, with transparency 4.1 s

The editor

A real editor around it.

Layers, masks, selections and adjustments, built so the generative tools fit in instead of taking over. Plain verbs, no sparkles.

Remove Background

One tap. A mask, not a deletion.

The background goes into a layer mask, so the original pixels stay underneath. An edge slider tunes the cut, and Select Subject refines it at full resolution.

A portrait of a woman in a teal sweater by a window. The same woman cut out onto a transparent background.

Upscale

Upscale on your own device.

Real-ESRGAN runs inside k-anvas, on your graphics card or your CPU, at 2x or 4x. It works offline and never touches your renderer.

A 4x upscale by classic resampling: the brass lamp stem is soft. The same crop upscaled 4x by k-anvas's on-device model: crisp edges and highlights. Classic k-anvas

Recipes

Every result remembers how it was made.

Prompt, seed, size, mask and model live with the layer. Reroll it, or edit the recipe and run it again, weeks later.

Fill · walnut basevariant 2 of 3
op
fill
prompt
Square dark walnut base
seed
1856410775
size
1024 × 1024 · Turbo
mask
marquee · feather 11 px
placed
outside the mask unchanged
Reroll6.6 s GPU

Layers

Keep picks as layers.

Every generation is a layer with a mask: hide it, paint its mask, change its blend mode, or bring back a variant you passed on.

  • SelectionsMarquee, lasso, magic wand and Select Subject, with add, subtract and feather.
  • AdjustmentsCurves, Levels and Hue/Saturation as adjustment layers, plus looks.
  • Branching historyUndo never throws work away. Go back, try something else, and both paths stay.
  • Command paletteEvery action a keystroke away, with familiar single-key tool shortcuts.
  • Brush, eraser, textPaint, retouch and set real type on its own layer.
  • LibraryEvery result goes to the Library first, searchable, with its recipe.

Your GPU

Your renderer. Your account.

k-anvas renders on an endpoint you deploy to your own Modal account. Your pictures go from your device to your GPU and back. There is no k-anvas server in between, no credits and no sign-up.

about 7 s
for a warm edit at one megapixel
65 to 130 s
to wake a cold GPU, shown as it happens, never a surprise
one token
per device, so each one can be paired and let go on its own
every result
shows its GPU seconds and what they cost you

Devices

At the desk and on the couch.

One app, laid out for the screen it is on. A mouse gets dense panels and single-key shortcuts. Touch gets 44-point targets, sheets and gestures.

k-anvas on desktop, with a generated lamp and the layers panel.
k-anvas on iPad, in landscape.
k-anvas on iPhone.
  • Desktop appmacOS, Windows and Linux
  • iPadAdd to Home Screen from Safari
  • iPhoneThe same app, laid out for one hand

Before launch: the desktop builds are not signed yet, and the iPad and iPhone layouts have been tested in browsers but not yet on real devices.

part of the k-family

  • k-holeOS
  • k-chaincompanion
  • k-flowagentic remote control
  • k-streammedia center
  • k-anvaspaint with a prompt.

Get k-anvas

The first release is almost here.

Desktop

macOS, Windows and Linux.

Coming soon

Web app

iPad, iPhone and desktop browsers. Installs to your Home Screen.

Coming soon

What you will need

A Modal account to run your renderer. k-anvas pairs with it in a minute, and you pay Modal only for the GPU seconds you use.

k-anvas renders with Qwen-Image-2.1, which is licensed under the Qwen Research License for research and evaluation only. Please don't use k-anvas for paid or client work.