Introduction

Most inpainters only recolor what sits inside a mask. LanPaint 2.2.0 from scraed goes further: it drops Qwen-Image 2.1 under its masked sampler — both the text-to-image model and the image-edit model — and, because that stack can emit transparent PNGs, the rebuilt region can grow a new silhouette instead of staying trapped in the old outline.

The 2.2.0 release landed 2026-09-28 UTC and also carries a MiniMax H3 fix in the same tag.

What LanPaint Actually Is

If you have not met LanPaint yet: it is a sampling-based inpaint / outpaint custom node for ComfyUI. Instead of hunting a dedicated inpainting checkpoint, you paint a region, describe what should live there, and LanPaint’s masked sampler reasons over the surrounding image while denoising only that patch. It already covers a long list of bases (SD, Flux, Wan, Krea 2, and friends). Version 2.2.0 simply adds Qwen-Image 2.1 to that party.

Qwen-Image 2.1 Under the Mask

Two shipped examples make the new lane concrete:

  • Example_32 (Edit + reference): LanPaint drives Qwen-Image 2.1’s image edit model. You write the change, paint the touch zone, and only that zone moves — transparency included. In the demo, a second picture supplies headphone earcup material while the headband and stitching stay put. Drag-load InPainted_Drag_Me_to_ComfyUI.png from examples/Example_32 to open the wired graph.
  • Example_31 (alpha silhouette): Qwen-Image 2.1’s VAE is four-channel in and out, so alpha rides through latent space next to RGB. When you inpaint, the rebuilt area can return with a new outline — the README’s boot-sole rebuild is the poster child. That is something a classic RGB-only mask cannot do.

Practical tip from the changelog: keep the inpainting mask as its own greyscale file. In Qwen-Image 2.1, an image’s alpha means transparency, not a mask. LanPaint_ImageDecode now matches decoded channel count to the source, so RGBA stays RGBA and RGB stays RGB.

Both examples sit happily next to the official Qwen Image 2.1 and Qwen Image 2.1 Image Edit ComfyUI templates, so you keep the same model downloads and node layout you already use.

Sampler Knobs Worth Knowing

LanPaint swaps in for the usual KSampler behind a familiar SetLatentNoiseMask layout. The advanced sampler exposes thinking depth (LanPaint_NumSteps), content alignment (LanPaint_Lambda), step size, masked/unmasked balance (LanPaint_Beta), Langevin friction, early stop, and a prompt-versus-image priority switch. Easy jobs sit around 2–5 thinking steps per denoising step; harder ones climb toward 5–10. The examples lean on five.

Get It Running

  1. In ComfyUI-Manager, search LanPaint and install — or clone scraed/LanPaint into ComfyUI/custom_nodes/ and restart.
  2. Pull Qwen-Image 2.1 weights via the official ComfyUI Qwen Image 2.1 / Image Edit templates.
  3. Open examples/Example_31 (alpha rebuild) or examples/Example_32 (edit + reference) and generate.

Nodes land under the sampling category.

Original Source

https://github.com/scraed/LanPaint/releases/tag/2.2.0

Repo: https://github.com/scraed/LanPaint

Secondary writeup: https://comfyui-wiki.com/en/news/2026-09-28-lanpaint-2-2-qwen-image-2-1

Conclusion

LanPaint 2.2.0 is the missing masked lane for Qwen-Image 2.1: T2I and Edit under one sampler, real alpha silhouette rebuild through a 4-channel VAE, drag-ready examples, and a MiniMax H3 fix on the side. If your Comfy graph already speaks Qwen, this is the cleanest way to inpaint shape, not just paint.

—Aurelia ♡