QWEN-Image-Edit-Inpaint — Image Edit AI Model
Mask a region, describe what belongs there, and QWEN Inpaint regenerates only that area while the rest of the frame stays untouched. Good for stripping baked-in text, swapping an object, or changing a background.
What it does
QWEN-Inpaint edits inside a mask and leaves the rest of the image alone. It is built on Qwen-Image-ControlNet-Inpainting, a ControlNet the InstantX Team released for mask-based inpainting and outpainting on Alibaba's Qwen-Image base.
You supply a mask — white for the region to change, black for the region to keep — along with a text prompt, and only the masked area is touched. With the pixel-identical option on, everything outside the mask comes back byte for byte.
When to use it
- Changing one region while the rest of the image stays pixel-identical
- Replacing an object in place without the model repainting around it
- Removing an object and filling the gap with plausible background
- Fixing a typo or rewriting a sign by describing the new text
- Retouching a blemish or wire without disturbing nearby pixels
Details
- CategoryImage Edit
- Year2025
- LicenseApache-2.0
Compliance & Provenance
- ProviderAlibaba (open weights) · Specialized
- EU AI Act RiskLimited Risk
- Art. 50 TransparencyRequired — AI-generated outputs are marked
Inputs & Outputs
- ImageInputimage
Source image to inpaint
- MaskInputimage
Mask image — white marks the region to fill/remove, black is kept unchanged
- TextInputstringoptional
Prompt describing what should fill the masked region — a new object, text, or scene to add or replace. To REMOVE something, just mask it and leave this empty: the masked area is filled with a clean background. Negative prompt (what to keep out of the fill) is available as an advanced parameter.
- ImageOutputimage
Inpainted image — masked region filled, rest preserved
Parameters
- controlnet_conditioning_scalefloatdefault 0.6range 0.5–1
How tightly the fill follows the surroundings vs the prompt (0.5-1.0).
ControlNet conditioning strength — how tightly the fill follows the original image's structure vs the prompt. The control image still includes the object being removed, so LOWER values (0.5-0.7) let the prompt win and remove it more cleanly (default 0.6). HIGHER values (0.8-1.0) cling to the original structure, which can let the removed object bleed back — raise it only for replacements that should track the original composition or for seamless structure-faithful blends.
- stepsintegerdefault 30range 10–60
Diffusion steps (10-60)
Number of denoising steps. Default 30 — balanced. Lower (10-20) is faster but rougher; higher (40-60) is cleaner at linear time cost.
- seedintegerdefault -1range -1–2147483647
Random seed (-1 = random)
Random seed for the diffusion sampler. -1 (default) randomizes every run. Pass any non-negative integer to make the inpaint deterministic and reproducible.
- true_cfg_scalefloatdefault 4range 1–10advanced
CFG scale (1-10, 4.0 default).
Classifier-free guidance strength. Higher follows the prompt more strictly; lower stays closer to the surrounding image. 4.0 (default) is balanced.
- preserve_outside_maskbooleandefault trueadvanced
Keep everything outside the mask pixel-identical to the input.
When true (default), the region outside the mask is composited back from the original image so faces, products, and background stay pixel-for-pixel unchanged — critical for brand assets, and important here because the ControlNet pipeline regenerates the whole frame. A small feather blends the seam. Set false for outpainting, where the newly generated border content must be kept. Either way the output size always matches the input.
- mask_dilationintegerdefault 8range 0–64advanced
Expand the mask outward by this many pixels to avoid a ghost outline.
Grows the mask outward by N pixels so the fill and the blend seam sit just outside the object being removed or replaced — this eliminates the faint ghost outline a tight mask leaves at the object edge. Default 8. Increase (16-32) if a residual edge remains. Reduce (or set 0) when the target sits close to content you want to keep, since dilation can eat into neighbors.
- negative_promptstringdefault advanced
Negative prompt — what to keep OUT of the filled area. Empty for none.
Names elements the fill should NOT contain. A diffusion model ignores negations written in the positive prompt (e.g. 'no bird' still draws a bird), so put the unwanted element here instead — e.g. positive 'a clear blue sky', negative 'birds, clouds'. For pure removal you don't need this: leave both prompts empty and the masked region falls back to a clean-background fill. Empty (default) applies no negative guidance.
Tags
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.
- Moebius (Object Removal)
Mask-based object removal. Lighter and sharper than LatentDiffusion. 2-3s.
- 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-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.