Say what to change in plain words, and Qwen Image 2.1 edits your picture without shifting it.<h3 id="why-use-this-workflow">Why use this workflow</h3>You can edit a picture with a plain Qwen Image 2.1 graph and a typed request. This workflow adds three things:<ul><li>It puts the edit back in place. Qwen draws an edit a little zoomed in or shifted, most of all on style changes. Realign to Source measures that and moves the edit back on top of your original. In my tests, style edits sat 35 px off at the median. After Realign: under 5 px.</li><li>It writes the instruction for you. Qwen3-VL looks at your picture and turns a short request into a full instruction. It changes only what you name.</li><li>It gives the edit room. A mirrored margin goes around your picture first, so the edit can move without leaving an empty strip at an edge.</li></ul><h3 id="two-versions">Two versions</h3><ul><li>v1.2 Fast does the edit in 6 steps with the Viggle Turbo LoRA. A run takes about 11 seconds on an RTX 5090. It needs one more file.</li><li>v1.1 does it in 25 steps. A run takes about 18 seconds. No extra file.</li></ul>The edit lines up the same in both.<h3 id="how-to-use-it">How to use it</h3><ol><li>Load your image.</li><li>Say the change in plain words, like "make it a soft watercolor painting".</li><li>Run.</li></ol><h3 id="what-you-need">What you need</h3><ul><li>AusBoss nodes 2.3.0 or newer. In ComfyUI-Manager, search "AusBoss". Already have them? Update first.</li><li>The Qwen Image 2.1 models, on a recent ComfyUI (I tested 0.37.0 and 0.38.0):<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">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">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">qwen_image_2.1_vae_bf16.safetensors</a> → <code>models/vae</code> · 676 MB</li></ul></li><li>For v1.2 Fast: <a target="_blank" rel="ugc" href="https://huggingface.co/Viggle/Qwen-Image-2.1-viggle-turbo/resolve/main/Qwen-Image-2.1-viggle-turbo-v0.3-6step-lora-r128.safetensors">Qwen-Image-2.1-viggle-turbo-v0.3-6step-lora-r128.safetensors</a> → <code>models/loras</code> · 680 MB</li><li>Optional: <a target="_blank" rel="ugc" href="https://huggingface.co/ausboss/Qwen-Image-2.1-Consistency-LoRA/resolve/main/qwen-image-2.1-consistency.safetensors">qwen-image-2.1-consistency.safetensors</a> → <code>models/loras</code> · 159 MB. Its row is off, so it runs without it.</li></ul><h3 id="tips">Tips</h3><ul><li>Say where a moved arm or hand ends up. "put her arm on the table" kept her clasped hands and added a third one. "put her arm on the table, her hands aren't clasped anymore" worked.</li><li>Check what Qwen got when an edit misses. Double-click Write the instruction to read the full instruction, then say the missing part more plainly.</li><li>The Consistency LoRA holds shapes on restyles. Turn its row on for those, and set Load Image + Pad's four pads to 0 inside Edit with Qwen Image 2.1. Leave it off for pose changes: it holds the old pose.</li><li>Realign fixes the frame, not redrawn details. Something the edit redrew in a new spot stays there.</li><li>In v1.2 Fast, leave the Viggle Turbo row on. Six steps without it come out soft and washed out.</li></ul><h3 id="credits">Credits</h3>Qwen Image 2.1 is by Alibaba's Qwen team, under the Qwen Research License: non-commercial use only. Comfy-Org repackaged the files. The speed LoRA in v1.2 Fast is <a target="_blank" rel="ugc" href="https://huggingface.co/Viggle/Qwen-Image-2.1-viggle-turbo">Viggle Turbo v0.3</a> by Viggle, under the same license: research use only. The Consistency LoRA is mine, under the same license.v1.2 Fast (2026-10-04): the same workflow with the edit in 6 steps, about 11 seconds a run instead of 18. v1.1 (2026-09-30): the same edit in a smaller, tidier workflow, and the prompt writer changes only what you ask for. v1.0 (2026-09-28): first release.Questions or bugs: <a target="_blank" rel="ugc" href="https://github.com/ausboss/ComfyUI-AusBoss/issues">GitHub issues</a> or the comments here. More from me 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>.

Qwen Image 2.1 Edit + Realign

<ul><li>The same workflow as v1.1 with one change: the Viggle Turbo LoRA does the edit in 6 steps instead of 25.</li><li>A run takes about 11 seconds on an RTX 5090 instead of about 18, with the models loaded.</li><li>The edit lines up the same. I ran 8 edits both ways with the same seed.</li><li>It needs one more file: Qwen-Image-2.1-viggle-turbo-v0.3-6step-lora-r128.safetensors, 680 MB, in models/loras. The run stops and names the file if it is missing.</li><li>v1.1 stays on the page for 25 steps with no extra download.</li><li>Needs AusBoss nodes 2.3.0 or newer, as before.</li></ul>

Files with SHA256: aca0840c5b12b84092a8a5e6006acc79ef2e22cafa7e347e770ba13ec6dd9a20