Moebius — Inpainting AI Model

What it does

Mark a region and Moebius fills it with content that continues the surrounding scene - a wire in a VFX plate, an unwanted reflection, an object sitting at the frame edge. Removal is the only task, so there is no prompt to write: two inputs and two sliders.

At 0.22B parameters it is small, which is the argument the model makes - a narrowly defined task does not need a generalist foundation model to do it well. Two to three seconds per edit, quick enough to adjust a mask and try again, and lighter and sharper than LatentDiffusion.

When to use it

• Removal at interactive speed, quick enough to iterate on the mask
• Wire and rig removal, set-dressing mistakes, frame-edge artefacts
• Nothing to describe, since removal is the only thing it does
• Batch cleanup across a large image set, at 0.22B parameters
• The rest of the frame left untouched, pixel for pixel

Details

• CategoryInpainting
• Year2026
• LicenseApache-2.0

Compliance & Provenance

• ProviderOpen-source (hustvl) · Specialized
• EU AI Act RiskLimited Risk
• Art. 50 TransparencyRequired — AI-generated outputs are marked

Inputs & Outputs

ImageInput · image

Source image with region to remove

ImageInput · image

Mask image marking the region to remove (white = remove)

ImageOutput · image

Output image with the masked region filled; outside the mask is unchanged

Parameters

seedinteger · default -1 · range -1–2147483647

Random seed (-1 = random)

Random seed for reproducibility. -1 (default) randomizes; fix to reproduce a previous result while iterating on the mask.

stepsinteger · default 20 · range 10–50

Diffusion steps (10-50)

Number of denoising steps. Default 20 — the value the authors report results with. Lower (10-15) is faster with softer fills; higher (30-50) rarely changes much on this model and costs time linearly.

guidance scalefloat · default 2.5 · range 1–7.5 · advanced

Guidance strength (1.0-7.5)

Classifier-free guidance strength. Default 2.5 (upstream default). Lower values stay closer to the surrounding background; higher values invent more. Runtime is unaffected — the model always evaluates both branches.

mask_dilationinteger · default 8 · range 0–64 · advanced

Expand the mask outward by this many pixels to avoid a ghost outline.

Expands the masked region outward before filling. A mask that traces an object exactly — which is what a Masking or Segmentation node produces — leaves the composite seam sitting on the object edge, so the seam re-draws a faint outline of whatever you removed. Measured on this model: without dilation the band just inside the mask edge changes only 0.73x as much as the mask interior, meaning the fill hugs the removed object; at the default 8 that ratio is 1.05, and it improved on all eight upstream sample pairs. Note the pixels are counted on the 512px canvas this model renders at, not on your source image — on a 2048px input, 8 here grows the mask by roughly 32 source pixels. Raise it (16-32) when a removal still shows an embossed shape; lower it toward 0 when the region butts up against detail you need kept sharp. 0 disables it.

Limitations

• Edited images are synthetic and have to be marked as such under the EU AI Act

Tags

• inpainting
• object-removal
• image-editing
• diffusion
• mask-based
• gpu-light
• fast

Alternatives in Image Edit

• FireRed-1.1

  Natural-language image edit with strong identity consistency. 1-3 min.

• LatentDiffusion (Object Removal)

  Mask-based object removal / inpainting. 10-30s.

• PowerPaint

  One mask, four jobs — remove, fill from a prompt, rebuild a border, or fill a shape.

• QWEN-2511

  Edits images from a natural-language instruction. High quality, and heavy (1–3 min).

• QWEN-Inpaint

  Prompt-driven mask inpaint — fills the masked area from your prompt, keeps the rest.

• QWEN-Layered

  Decomposes an image into editable RGBA layers (foreground objects, background, …). Pair with List-Extract → QWEN-Image-Edit → List-Inject → Layer-Compose to edit a single layer and recompose.

Resources

• github
• paper