Image Edit - Moebius (Object Removal)

Image Edit - Moebius (Object Removal)

🛠

Mask-Based Object Removal — 0.22B Parameters, 2 – 3 Seconds

Moebius is a lightweight image inpainting model from HUST Vision Lab and VIVO AI Lab (ECCV 2026). Give it an image and a mask; it removes what's inside the mask and rebuilds the background. No prompt, no options to reason about — the model conditions on a learned embedding table rather than a text encoder, so there is no text input at all.

The reason it's interesting is the size. At 0.22B parameters it is roughly 2% the size of FLUX.1-Fill-Dev (11.9B), and the authors report matching or beating that model's generation quality across six benchmarks while running more than 15× faster end to end. In CNAPS Studio a removal lands in 2 – 3 seconds.

Weights MIT · code Apache 2.0. Both permissive.

What it does

One job, done fast. Mark a region, and Moebius fills it with content that continues the surrounding scene — the wire in a VFX plate, the tourist in the corner of a shot, the reflection you didn't want, the object at the frame edge.

What makes it worth a node of its own is a deliberate argument about model scale. Inpainting quality has been climbing by making generalist foundation models bigger, which puts good inpainting behind serious compute. Moebius takes the opposite position: if the task is narrowly defined, a specialist can be smaller, faster and no worse. The authors compress to 0.22B by summarising spatial context and global semantic priors into fixed-size linear matrices instead of carrying them in parameters, then recover the quality that compression would normally cost through distillation carried out entirely in latent space — no expensive decode back to pixels during training.

The practical result is a removal model fast enough that removing something stops feeling like a job you queue and starts feeling like a click.

Problem it solves

  • Removal at interactive speed — 2 – 3 seconds per edit. Fast enough to try a mask, look at it, adjust the mask and try again, rather than batching removals and reviewing them later
  • No prompt to write — Removal is the only task, so there's nothing to describe. Two inputs, two sliders
  • Cheap enough to run at volume — 0.22B is a small model by any current standard. Batch removal across a large image set is a reasonable thing to do here in a way it isn't with a 10B-class editor
  • Quality that doesn't obviously trade down for size — The authors' benchmark claim is parity with a model roughly 50× larger. That claim is theirs and worth verifying on your own footage, but it is the point of the model
  • A clean licensing position — MIT weights, Apache 2.0 code, no territorial or revenue conditions
  • Fits VFX and cleanup workflows — Wire and rig removal, set-dressing mistakes, unwanted reflections and frame-edge artefacts are the classic uses for exactly this operation

Input/Output

  • Input:two ports, both required (filled connection points)
    • Image — the source image
    • Mask — the region to remove and rebuild

      • There is no text input, and that isn't an omission. Moebius conditions on a learned embedding table instead of a text encoder, so the architecture has nowhere for a prompt to go. If you need to describe what should appear in the masked area rather than have it disappear, use PowerPaint or QWEN-Inpaint instead. Port identity — which of the two is image and which is mask — is inferred from port order and should be confirmed against the node before it goes in a tutorial

  • Output: a single image with the masked region rebuilt
    • No mask, confidence map, or intermediate is returned
    • The model is a removal specialist; there is no mode selector
  • Parameters: four in total, two on the node and two under Advanced.
    • Random seed (1 = random) — fix it to reproduce a result. Worth using: diffusion removal is non-deterministic, and if a mask produces an artefact, re-running the same mask on a new seed is often the whole fix
    • Diffusion steps (slider, 10 – 50, default 20) — lower than the usual 30 for this class of model, consistent with the model being built for speed
    • Guidance strength (1.0 – 7.5, default 2.5) — note the range. Prompt-driven inpainting models typically run 7.5 and up; Moebius tops out at 7.5 and defaults to 2.5, because guidance here steers a learned embedding rather than a text prompt. Don't carry settings over from a text-guided node
    • Expand the mask outward by this many pixels (default 8) — dilates the mask so no rim of the original object survives at the boundary. If you can still see a faint outline of what you removed, raise this before touching anything else

Performance

The figures below are the authors' published claims for the underlying model unless marked otherwise. They have not been independently re-verified against this deployment.

Parameters
0.22B
Reference comparison
Authors report rivalling or surpassing FLUX.1-Fill-Dev (11.9B) across six benchmarks spanning natural and portrait scenes — under 2% of the parameters
Speed (upstream claim)
More than 15× faster in total inference time than the 11.9B baseline
Speed (this node)
2 – 3 seconds per removal
Diffusion steps
10 – 50, default 20
Sampler
DDIM with classifier-free guidance (reported by a third-party ONNX port, not the official repo)
Native resolution
The model's cross-attention uses a position embedding tied to its trained resolution of 512 × 512; a third-party port letterboxes non-square input. How the CNAPS node handles resolution and aspect ratio is not documented here — verify on your own asset sizes before a batch run
Lineage
Direct continuation of the same lab's PixelHacker
End-to-end node timing at scale
The 2 – 3 s figure is per removal. Batch throughput has not been measured
‣
Technical Details (for engineers)

How Moebius compares

CNAPS Studio has four ways to edit inside a mask. They separate cleanly by what you're trying to do.

Moebius
LatentDiffusion (Object Removal)
PowerPaint
QWEN-Inpaint
Remove an object
✅
✅
✅
✅
Generate from a prompt
❌ (no text input)
❌
✅
✅
Outpaint / shape-guided
❌
❌
✅
❌
Parameter class
0.22B
Latent diffusion
SD 1.5-class
Qwen-Image-Edit-2511
Speed
2 – 3 s
—
—
—

Rough guidance: Moebius when the job is removal and you want it fast — it is the narrowest and lightest of the four, and the node description positions it as lighter and sharper than the LatentDiffusion removal node. PowerPaint when one node needs to cover several editing intents. A larger editor when single-image fidelity outweighs throughput.

Speed figures for the other three haven't been measured; only Moebius carries a published number. Scope claims come from each node's own documentation, and the models have not been benchmarked head-to-head inside CNAPS Studio.

All of these run inside CNAPS Studio — compare them side-by-side.

Model Source

Compliance & Provenance

Provider
HUST Vision Lab (Huazhong University of Science and Technology) with VIVO AI Lab
Provider type
Specialized
License
Model weights MIT as listed on the Hugging Face model card; reference code Apache 2.0. Both permit commercial use with no territorial or revenue conditions. Note: permissive licensing covers the weights and code — it is not a statement about training-data provenance, and it is not IP indemnification
EU AI Act risk class
Limited Risk
Art. 50 transparency
Required — output is AI-modified image content and is marked with a visible AI label and machine-readable provenance. See AI Policy §2.
Region availability
Available globally
Training data summary
Not published as a structured summary per Art. 53(d); see the paper for training methodology

For more on how we classify models and mark outputs, see our AI Policy.

Run it now