Turn, crop and open up a picture with Qwen Image 2.1 and my outpaint LoRA, in one pass. Straighten or tilt it, cut away what you don't want, and pad the sides out. Qwen paints the new areas, and your kept pixels go back on top untouched.The first run turns the sample mirror selfie by -14.2°, cuts away the mirror frame, and opens the bedroom up on the left, top and right: 864 × 1152 in, 864 × 1216 out, in about 8 seconds on an RTX 5090.<h2 id="what-it's-for">What it's for</h2><ul><li>Straighten a crooked photo without losing its corners: the rotation's gray wedges get painted.</li><li>Cut something away and let it be painted fresh, like the mirror frame in the sample.</li><li>Re-roll a part you don't like: crop past it, then pad back out.</li><li>Put a transparent cutout in a scene: its transparent area is painted too. Say which scene in the extra-detail line.</li></ul><h2 id="what's-inside">What's inside</h2><ul><li>Image Crop + Rotate + Pad: turn the picture with the round handle, drag the crop edges in, and drag the orange handles out. Flat gray (#808080) fill, feather 0, and a 1 MP canvas in steps of 32. Load your picture first: picking a new one resets the tilt, crop and padding.</li><li>Auto prompt: Text Generate runs the Qwen3-VL 8B encoder that Qwen 2.1 already loads (no extra download). It describes the canvas Qwen gets, so anything you cropped away never reaches the prompt. The prompt is the outpaint instruction, then <code>Scene:</code>, then that description; Show Text displays it.</li><li>Extra detail (optional): one line for what the new area should hold. The sample's line asks for more of the bedroom.</li><li>Encode + sample: Text Encode Qwen Image 2.1 takes the canvas as <code>image_1</code> at resolution 0, then 25 steps, CFG 1, <code>euler</code> / <code>simple</code>, fixed seed. My outpaint LoRA v2 runs at strength 1, with v1 one click away. If the LoRA file is missing, the run stops before sampling and the LoRA Loader turns red. Switch the row off to run plain Qwen 2.1.</li><li>Stitch + save: Split Image with Alpha drops the alpha channel that Qwen 2.1's VAE decodes. Stitch Inpaint puts your kept pixels back with a 32 px blend, Compare slides the result against the canvas, and Save Image writes a PNG with the workflow embedded.</li></ul><h2 id="quick-start">Quick start</h2><ol><li>Update ComfyUI. You need a build with Text Encode Qwen Image 2.1 (mid-September 2026 or newer; I tested on 0.37.0).</li><li>Install AusBoss nodes from ComfyUI-Manager (search "AusBoss", registry id <code>ausboss-nodes</code>) and restart. You need 2.2.0 or newer, which is what Manager installs.</li><li>Unzip <code>sample_inputs.zip</code> into <code>ComfyUI/input/</code>.</li><li>Drag the workflow in and download the files from the card, the LoRA included.</li><li>Queue. For your own picture, load it first, then turn, crop and pad.</li></ol><h2 id="models">Models</h2><ul><li><a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/Qwen-Image-2.1/resolve/main/diffusion_models/qwen_image_2.1_int8_convrot.safetensors?download=true">qwen_image_2.1_int8_convrot.safetensors</a> → <code>models/diffusion_models</code> · 7.26 GB</li><li><a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/Qwen-Image-2.1/resolve/main/text_encoders/qwen3vl_8b_int8_convrot.safetensors?download=true">qwen3vl_8b_int8_convrot.safetensors</a> → <code>models/text_encoders</code> · 9.35 GB</li><li><a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/Qwen-Image-2.1/resolve/main/vae/qwen_image_2.1_vae_bf16.safetensors?download=true">qwen_image_2.1_vae_bf16.safetensors</a> → <code>models/vae</code> · 676 MB</li><li>recommended <a target="_blank" rel="ugc" href="https://huggingface.co/ausboss/Qwen-Image-2.1-Outpaint-LoRA/resolve/main/qwen-image-2.1-outpaint-v2.safetensors?download=true">qwen-image-2.1-outpaint-v2.safetensors</a> → <code>models/loras</code> · 159 MB</li><li>or <a target="_blank" rel="ugc" href="https://huggingface.co/ausboss/Qwen-Image-2.1-Outpaint-LoRA/resolve/main/qwen-image-2.1-outpaint.safetensors?download=true">qwen-image-2.1-outpaint.safetensors</a> (v1, big zoom-outs) → <code>models/loras</code> · 159 MB</li></ul><h2 id="custom-nodes">Custom nodes</h2>Only <a target="_blank" rel="ugc" href="https://github.com/ausboss/ComfyUI-AusBoss">ComfyUI-AusBoss</a>, 2.2.0 or newer (Manager installs it): free and open source. Text Encode Qwen Image 2.1, the reference cache, Text Generate and Split Image with Alpha are core ComfyUI.Manager can't find AusBoss? In the classic Manager window, if the Channel box is empty, pick default, then click Install Missing Custom Nodes again. <a target="_blank" rel="ugc" href="https://github.com/ausboss/ComfyUI-AusBoss#if-manager-cant-find-ausboss">See what to click</a>.<h2 id="settings-that-worked-and-what-didn't">Settings that worked, and what didn't</h2><ul><li>Use the LoRA. On tilted pictures, plain Qwen 2.1 moved the scene in all 12 of my test renders, by up to 95 px. The pasted-back original then showed as a lighter tilted box with doubled edges. With the LoRA, the picture stayed put in all 24.</li><li>Feather 0. This node's feather also fades the picture into the gray, and the model paints that fade as a darker band along the tilt. Stitch Inpaint still blends the seam.</li><li>Color match 0. On a rotated and padded canvas it reads the gray corners as part of the picture and tints the fill.</li><li>The caption reads the canvas, not your file. Your file can still hold what you cropped away, or black under transparent pixels, and describing those brought them back. Reading the canvas, it invents a little more (it once called a record crate a piano); the extra-detail line steers it.</li></ul><h2 id="speed">Speed</h2>Measured on an RTX 5090 (32 GB) with the models loaded: about 8 s for the sample, caption included.<h2 id="about-the-gallery">About the gallery</h2>The PNG is a straight output of this workflow from the sample; drag it into ComfyUI to load its exact settings. The sample is my own render. The cover is cut from the real run.<h2 id="license-and-credits">License &amp; credits</h2>Qwen Image 2.1 is © Alibaba Qwen, released under the Qwen Research License: non-commercial use only; commercial use needs a separate license from Qwen. The outpaint LoRA is mine and follows that license when used with the model. The workflow and the AusBoss nodes are free.<h2 id="changelog">Changelog</h2><ul><li>v1.0 (2026-09-26): first release.</li></ul>Questions or bugs: <a target="_blank" rel="ugc" href="https://github.com/ausboss/ComfyUI-AusBoss/issues">GitHub issues</a> or the comments here. Post what you make with it; I read everything. I post new workflows on <a target="_blank" rel="ugc" href="https://x.com/Zanzibased">X @Zanzibased</a> and <a target="_blank" rel="ugc" href="https://github.com/ausboss">GitHub</a>.

Files with SHA256: f68d1d4e13b25d124c95f68f2c3636d4e5766497c9a6acbf72fe4e49966d9018