Upload a picture and frame the new area on the interactive outpaint canvas (a custom gr.HTML
component modeled on the Krea-style canvas of
yijunwang2/krea2-outpaint): drag your picture
to position it, drag the canvas edges / corners to extend it, or snap to ratio and anchor presets. The
hatched gray extension area is the mask that the
Qwen-Image-2.1 Outpaint LoRA by ausboss
fills in on Qwen/Qwen-Image-2.1 while keeping
the original picture pixel-registered.
It follows the author's ComfyUI recipe: gray #808080 padding on a /32 canvas, the padded canvas as
the reference image at the same output size, the LoRA's instruction prompt plus an automatic
Scene: caption written by Qwen-Image-2.1's own Qwen3-VL text encoder, 25 steps at CFG 1, and the
original pixels pasted back over a 32 px feather. Both LoRA versions (v1, v2) and a no-LoRA baseline
can be picked under Advanced.
Base model and LoRA are under the Qwen Research License (non-commercial). Example photos are CC0, from linoyts/repo-to-space-example-inputs.
Upload a picture and frame the new area on the interactive outpaint canvas (a custom gr.HTML
component modeled on the Krea-style canvas of
yijunwang2/krea2-outpaint): drag your picture
to position it, drag the canvas edges / corners to extend it, or snap to ratio and anchor presets. The
hatched gray extension area is the mask that the
Qwen-Image-2.1 Outpaint LoRA by ausboss
fills in on Qwen/Qwen-Image-2.1 while keeping
the original picture pixel-registered.
It follows the author's ComfyUI recipe: gray #808080 padding on a /32 canvas, the padded canvas as
the reference image at the same output size, the LoRA's instruction prompt plus an automatic
Scene: caption written by Qwen-Image-2.1's own Qwen3-VL text encoder, 25 steps at CFG 1, and the
original pixels pasted back over a 32 px feather. Both LoRA versions (v1, v2) and a no-LoRA baseline
can be picked under Advanced.
Base model and LoRA are under the Qwen Research License (non-commercial). Example photos are CC0, from linoyts/repo-to-space-example-inputs.