Load a picture and set the megapixels. Qwen Image 2.1 redraws it sharp in tiles, to 4 or 8 megapixels. Nothing to type.<h3 id="why-use-this-workflow">Why use this workflow</h3>Use it to take a phone-size photo, an old save or a soft JPEG to 4 or 8 megapixels.<ul><li>It redraws your picture instead of stretching it. A plain upscale model made my test pictures bigger without adding real detail. Qwen Image 2.1 looks at the picture and draws it again, sharp.</li><li>One number sets the size. Up to about 2 megapixels is one pass. Above that the picture is cut into tiles of about 2 megapixels, each tile is redrawn, and the tiles are joined. 4 megapixels takes two tiles, 8 takes four.</li><li>No line where the tiles meet. Neighbouring tiles share a strip and fade into each other. I looked for the joins in my results and did not find one.</li><li>It stays your picture. It runs with my <a target="_blank" rel="ugc" href="https://civarchive.com/models/2989697/qwen-image-21-restore-lora">Restore LoRA</a>, which keeps the redraw in place. Its page has a plain one-pass workflow for 2 megapixels, and its <a target="_blank" rel="ugc" href="https://huggingface.co/ausboss/Qwen-Image-2.1-Restore-LoRA">Hugging Face card</a> has the numbers.</li><li>It keeps your colors, also on a noisy photo. Each tile is matched to your picture before the tiles are joined. Grain and color noise do not push the saturation up.</li><li>You see what changed. A slider compares before and after.</li></ul><h3 id="how-to-use-it">How to use it</h3><ol><li>Load your image. Optional: drag the blue handles to crop it first.</li><li>Set Megapixels: the size of the result.</li><li>Run.</li><li>Slide Before and after to compare. The result is saved as a PNG.</li></ol>The request Qwen gets is already in the text box. It works as it is.<h3 id="megapixels">Megapixels</h3><ul><li>Up to about 2: one pass.</li><li>Over that: tiles of about 2 megapixels, joined.</li><li>A small picture is not pushed far past what it holds. It is redrawn at up to 8 times its own megapixels and enlarged the plain way from there. A tiny picture redrawn at 8 megapixels gets a made-up face.</li><li>A picture of 1 megapixel or more has to grow to look better, about 1.6 times per side. A 1 megapixel photo at 2 looks the same; at 4 or 8 it gains.</li><li>A big photo that is noisy or soft: set 2 first, run, then load the result and set the size you want. Cut at its own size, its tiles can come back grainy.</li></ul><h3 id="fast">Fast</h3>The Fast switch is on: 8 steps with the Viggle Turbo LoRA. Click it off for 25 steps without that LoRA. That takes about three times as long. In my measurements it came out a little closer to the picture.<h3 id="what-you-need">What you need</h3><ul><li>AusBoss nodes 2.9.0 or newer. In ComfyUI-Manager, search "AusBoss". Already have them? Update first. The two nodes that do the tiles, Tiled Upscale and Tiled Upscale Stitch, are new in 2.9.0.</li><li>ComfyUI 0.38 or newer.</li><li>The Qwen Image 2.1 models:<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>My Restore LoRA: <a target="_blank" rel="ugc" href="https://huggingface.co/ausboss/Qwen-Image-2.1-Restore-LoRA/resolve/main/qwen-image-2.1-restore.safetensors">qwen-image-2.1-restore.safetensors</a> → <code>models/loras</code> · 159 MB</li><li>For 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. Switch Fast off and the workflow runs without it.</li></ul><h3 id="tips">Tips</h3><ul><li>Blurry photo (out of focus, shaky, a frame from a video)? Set the Restore LoRA to 0.5. It is the first row in the LoRA Loader. At 1.0 it stays true to the blur and the result comes back soft. Pictures that are in focus, also small, noisy or blocky ones, do best at 1.0.</li><li>Made-up detail changes with the seed. If a small thing bothers you, run it again.</li><li>Keep colors is inside the box, on the Tiled Upscale Stitch node. At 1 each tile holds your picture's colors. Lower it and Qwen's own colors come through.</li><li>Zooming in a lot? Load the Texture-Fix VAE in place of the normal one: <a target="_blank" rel="ugc" href="https://huggingface.co/madebyollin/texture-fix-vae-for-qwen-image-2.1/resolve/main/texture_fix_vae_for_qwen_image_2.1_bf16.safetensors">texture_fix_vae_for_qwen_image_2.1_bf16.safetensors</a> → <code>models/vae</code> · 676 MB, by <a target="_blank" rel="ugc" href="https://huggingface.co/madebyollin/texture-fix-vae-for-qwen-image-2.1">madebyollin</a>. It removes a fine grid the normal VAE leaves. You only see the difference at about 400%.</li><li>If the run stops at a red LoRA Loader, a LoRA file is missing. Download it from the list above.</li><li>The two nodes work with other models too. Anything between Tiled Upscale and Tiled Upscale Stitch runs once per tile.</li><li>On an RTX 5090 with Fast on: about 9 seconds for one pass, about 20 for 4 megapixels in two tiles, about 43 for 8 megapixels in four.</li></ul><h3 id="what-it-does-not-do">What it does not do</h3><ul><li>It does not bring back what the camera saw. It is a redraw, and fine detail is made up to fit. A face stays the same person, not the same pixels.</li><li>A very small or very blurry face can change. It comes back sharp and believable, and it can look less like the person.</li><li>Small lettering can come back misspelled. Signs, labels and book spines. It differs from seed to seed.</li><li>A sharp picture that is already big gains little. It has to grow to look better.</li><li>A blurry photo needs the LoRA lower. Out of focus, shaky or a frame from a video: at 1.0 the LoRA stays true to the blur and the result comes back soft. Set the Restore row to 0.5 (see Tips).</li><li>Everything was measured on one RTX 5090 with one seed per picture.</li></ul>About the examples: they are my own sharp pictures at about 1 megapixel, one at half a megapixel. I made each one worse first, the way small pictures usually are: noise, a heavy JPEG save, or a soft and noisy JPEG. The workflow got only that file. Six are at 8 megapixels in four tiles, one at 4 in two. None of them is in the LoRA's training set.<h3 id="credits">Credits</h3>Qwen Image 2.1 is by Alibaba Qwen, under the Qwen Research License: non-commercial use only. The speed LoRA 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 Qwen Research License: research use only.v1.0 (2026-10-07): 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 Tiled Upscale

First release. Needs ComfyUI 0.38 or newer and AusBoss nodes 2.9.0 or newer: already have them? Update them in ComfyUI-Manager first, then restart ComfyUI.

Files with SHA256: 88ba0504f9805516513eaaa565215f402ab46365af61513916dd7315f55cf5a6