Qwen Inpainting Sampler

Qwen-Image-Edit is great at instruction edits but has one structural weakness: it re-emits the whole frame, so pixels you never asked to touch come back slightly different, and the drift compounds. The fix the community keeps landing on is to put a mask back around it and only let the model change the region you painted. QwenInpaintSampler is a diffusers-pattern inpainting sampler that does exactly that - you give it a latent, a mask, and a strength, and it confines the edit to the masked area.

The pack lists this among its experimental nodes, so treat it as capable-but-rough rather than a polished production sampler. Still, for the very common "edit just his shirt, leave his face alone" job, having a real masked sampler beats letting a full-frame edit wander.

How it works

It follows the diffusers inpainting approach: the masked region gets denoised toward your prompt while the unmasked region is anchored to the original latent, with a strength control governing how hard the change is pushed. That's the important lever - at strength 1 the masked area is fully regenerated; lower it and you blend the new content with what was there, which is how you keep an edit from looking pasted in. It defaults to 8 steps and true_cfg_scale 1, which reads like it's tuned for the fast, step-reduced (Lightning-LoRA) way most people actually run the 20B Qwen model.

There's also a padding_mask_crop option, which crops around the mask before sampling so the model works on a tight region at full detail rather than the whole image - the crop-and-stitch idea, built in.

The inputs and outputs that matter

• model, positive, negative - your Qwen model and the conditioning from a Qwen encoder.
• latent_image (required, LATENT) - the image you're inpainting, as a latent.
• mask (required, MASK) - the region to change. This is the whole point; everything outside it is protected.
• strength (default 1.0) - how completely the masked area is regenerated. Drop it to blend rather than fully replace.
• steps (default 8) and true_cfg_scale (default 1.0) - low defaults suited to step-reduced Qwen setups; raise them if you're running the model at full steps.
• padding_mask_crop (default 0, optional) - crop around the mask for detail; set a pixel padding to enable it.

Output: a single LATENT to hand to your VAE decode.

How to install it

ComfyUI Manager: search ComfyUI-QwenImageWanBridge, install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/fblissjr/ComfyUI-QwenImageWanBridge

then restart. Needs a Qwen-Image-Edit checkpoint and the Qwen2.5-VL encoder in the graph; no model of its own.

Common issues & troubleshooting

The edit bleeds outside the mask. Check that your mask is clean and actually covers only what you want changed. Feathering the mask edges helps the transition, but a sloppy mask is the usual culprit for spill.

The inpaint looks pasted-in or too aggressive. Lower strength so the new content blends with the original rather than fully overwriting it, and consider padding_mask_crop so the model sees context around the region.

It's experimental - expect rough edges. The pack files this under its experimental nodes. If it misbehaves and you just need reliable masked editing, the mainstream route is Qwen-Image-Edit with a native crop-and-stitch inpaint workflow. This node is the pack's own take on it, not the battle-tested path.