The image API, on one page

flux-klein — Generative edit, upscale and restore

The general-purpose editor: your prompt decides what happens, and the output is the size you ask for.

Capability generative_repaint · verified 2026-09-26 (public gateway) · JSON for this model

References

1. image_ids[0] — the image to edit

Returns: PNG or WebP at the width and height you sent; 1–4 references may be combined.

Measured: ~16 s end to end for a 1152 × 1536 render, 6.4 s of it inference (measured 2026-09-26); the rest is queueing and the result upload

Parameters

Where it stops

• It is a generative re-render: faces and small text can shift. Review the result at 100% before publishing.
• It cannot add detail that was never captured with certainty — an upscale reconstructs plausible texture, not the original pixels.

{
  "model": "flux-klein",
  "image_ids": [
    "<file_id>"
  ],
  "prompt": "Restore realistic detail, texture and sharpness at 2x the size while keeping the composition identical. Do not add objects and do not change the framing.",
  "width": 1152,
  "height": 1536,
  "num_inference_steps": 8,
  "seed": 42,
  "output_format": "png"
}

Open the enhancer in the studio

pixrestore-s — Detail restore

Renders a sharper, cleaner version of an image at its own size. No prompt, no knobs beyond steps.

Capability detail_restore · verified 2026-09-17 (public gateway, 8/8 with the idempotency pair) · JSON for this model

References

1. image_ids[0] — the image

Returns: PNG at the input size.

Measured: ~3.5 s for a 3131 × 1549 frame (measured 2026-09-17, end to end)

Parameters

Refused, not ignored: width, height, max_sequence_length, output_format, output_quality — sending one fails the call instead of being dropped.

Where it stops

• The result is "clearer", not "faithful": detail is generated, so fine text and small patterns can change. It is not a lossless enlarger.
• It does not resize: send the size you want to keep, or upscale separately.

{
  "model": "pixrestore-s",
  "image_ids": [
    "<file_id>"
  ],
  "num_inference_steps": 1,
  "seed": 42
}

cutout-fast — Background removal

Cuts the subject out and answers a PNG with an alpha channel at the input size.

Capability tool_cutout · verified 2026-09-26 (public gateway) · JSON for this model

References

1. image_ids[0] — the photo

Returns: PNG with transparency, at the input size — 768 × 1024 in, 768 × 1024 out.

Measured: ~5 s end to end, 0.8 s of it inference (measured 2026-09-26)

Parameters

Refused, not ignored: width, height, max_sequence_length, output_format, output_quality — sending one fails the call instead of being dropped.

Where it stops

• Fine hair, glass and motion blur keep some background; a see-through subject cannot be cut cleanly.
• The cut is the whole frame. There is no per-object selection: one subject per call.

{
  "model": "cutout-fast",
  "image_ids": [
    "<file_id>"
  ]
}

Remove a background in the studio

erase-quality — Generative erase

Names what should go; the model removes it and rebuilds the surface underneath.

Capability tool_erase · verified 2026-09-22 (public gateway, 8/8) · JSON for this model

References

1. image_ids[0] — the photo

Returns: PNG at a fixed 1024 × 1024, whatever the input size was.

Measured: ~54 s end to end, 50 s of it inference (measured 2026-09-22)

Parameters

Refused, not ignored: width, height, max_sequence_length, output_format, output_quality — sending one fails the call instead of being dropped.

Where it stops

• The whole frame is re-rendered at 1024², so a larger original comes back at that size and fine text elsewhere can drift.
• It removes what you name; it does not understand "remove the person but keep the shadow".

{
  "model": "erase-quality",
  "image_ids": [
    "<file_id>"
  ],
  "prompt": "Remove the spoon from the table and rebuild the wooden surface behind it; keep the mug, the napkin and the lighting exactly as they are.",
  "num_inference_steps": 8,
  "seed": 42
}

Erase an object in the studio

vectorize-fast — Image to SVG

Traces a bitmap into real vector paths and answers an SVG — so it scales to any size without blurring.

Capability tool_vectorize · verified 2026-09-22 (public gateway, 8/8) · JSON for this model

References

1. image_ids[0] — the bitmap

Returns: image/svg+xml. The SVG declares its own width and height, so the JSON result reports width/height as null.

Measured: ~7 s end to end, 0.4 s of it tracing (measured 2026-09-22)

Parameters

Refused, not ignored: width, height, max_sequence_length, output_format, output_quality — sending one fails the call instead of being dropped.

Where it stops

• A photograph comes back posterised: the photo preset limits the palette to 16 colours before tracing.
• The output is paths, not an embedded bitmap — expect a small file for a logo and a large one for a picture.

{
  "model": "vectorize-fast",
  "image_ids": [
    "<file_id>"
  ],
  "preset": "logo"
}

Convert an image in the studio

tryon-quality — Virtual try-on

Puts the garment from one photo onto the person in another. Two references, in order, and no prompt.

Capability tool_tryon · verified 2026-09-26 (public gateway, re-measured); the 10/10 pass including the idempotency pair was 2026-09-25 · JSON for this model

References

1. image_ids[0] — person
2. image_ids[1] — garment

Returns: PNG shaped like the person photo (3:4 → 768 × 1024). Face, hair, pose, other clothing and background are asked to stay identical.

Measured: ~69 s end to end, 59 s of it inference (measured 2026-09-26)

Parameters

Refused, not ignored: prompt, width, height, num_inference_steps, max_sequence_length, output_format, output_quality — sending one fails the call instead of being dropped.

Where it stops

• One garment per call: the reference is read as a single piece, so a top and a bottom need two calls.
• The order of image_ids is the meaning — person first, garment second. Swapping them changes the result, not the error.
• Layered outfits, heavy shadows across the garment and unusual poses are the hard cases.

{
  "model": "tryon-quality",
  "image_ids": [
    "<person_file_id>",
    "<garment_file_id>"
  ],
  "garment_type": "top"
}

Try it on in the studio

interior-quality — Interior render

Turns an empty or unfinished room into a furnished render in a named style, keeping the architecture where it is.

Capability tool_interior · verified 2026-09-25 (public gateway, 8/8) · JSON for this model

References

1. image_ids[0] — room
2. image_ids[1] — style reference (optional)

Returns: PNG shaped like the room photo (3:2 → 1152 × 768), with walls, windows, doors, ceiling line and camera angle held.

Measured: ~59 s of inference (measured 2026-09-25)

Parameters

Refused, not ignored: prompt, width, height, num_inference_steps, max_sequence_length, output_format, output_quality — sending one fails the call instead of being dropped.

Where it stops

• Start from the emptiest frame you have: a room that is already furnished comes back redesigned rather than built on.
• The second reference steers palette and materials only — it cannot add a wall or move a window.
• It is a visualisation, not a measured plan: sockets, floor coverings and furniture proportions can differ from the building.

{
  "model": "interior-quality",
  "image_ids": [
    "<room_file_id>"
  ],
  "style": "nordic",
  "room_type": "living_room",
  "extra": "light oak flooring, a linen sofa, a round wooden coffee table, a woven rug, a floor lamp and green plants"
}

Render a room in the studio

retouch-quality — Portrait retouch

Cleans up a portrait — skin, under-eyes, eyes, stray hairs — without changing the face.

Capability tool_retouch · verified 2026-09-25 (public gateway, 8/8) · JSON for this model

References

1. image_ids[0] — portrait

Returns: PNG shaped like the portrait, with skin grain kept.

Measured: ~79 s end to end, 61 s of it inference (measured 2026-09-25)

Parameters

Refused, not ignored: prompt, width, height, num_inference_steps, max_sequence_length, output_format, output_quality — sending one fails the call instead of being dropped.

Where it stops

• It will not slim a face, widen the eyes or change a hairstyle: that is the instruction, not a setting.
• It does not add makeup — that is the separate virtual-makeup model.
• A heavy filter look is refused on purpose; the result keeps visible texture.

{
  "model": "retouch-quality",
  "image_ids": [
    "<portrait_file_id>"
  ],
  "level": "natural"
}

Retouch a portrait in the studio

makeup-quality — Virtual makeup

Applies a named makeup look to a face, optionally matched from a reference photo.

Capability tool_makeup · verified 2026-09-25 (public gateway, 8/8) · JSON for this model

References

1. image_ids[0] — portrait
2. image_ids[1] — makeup reference (optional)

Returns: PNG shaped like the portrait. Identity, expression, hair, clothing, framing and lighting are asked to stay as they are.

Measured: ~62 s of inference (measured 2026-09-25)

Parameters

Refused, not ignored: prompt, width, height, num_inference_steps, max_sequence_length, output_format, output_quality — sending one fails the call instead of being dropped.

Where it stops

• A look is a recipe, not a brand shade: use it to compare directions, not to promise an exact product.
• The palette is matched to this person’s undertone and the original lighting, so the same reference gives a different result on another photo.

{
  "model": "makeup-quality",
  "image_ids": [
    "<portrait_file_id>"
  ],
  "look": "korean",
  "intensity": "medium"
}

Try a look in the studio