Make a picture from a prompt, or change one you load, with the official Qwen Image 2.1 Turbo checkpoint. 8 steps, about 4 seconds for 2 megapixels on an RTX 5090. One switch writes the prompt for you, one upscales the result.

How to use it

1. Write your prompt. With Auto prompt on, a few words are enough.

2. Set the size in Latent Size.

3. Run.

4. To change a picture instead, switch Edit a picture on, load your picture and write the change as the prompt.

The three switches

• Edit a picture. Off: the prompt makes a new picture. On: the prompt changes the picture you load, at 2 megapixels in your picture's own shape. It also turns my Consistency LoRA on, which keeps the edit lined up with your picture.

• Auto prompt. Qwen writes the full prompt from a few words of yours. For an edit it looks at your picture and writes the instruction. What it wrote shows in Show Text. It uses the text encoder the workflow already loads, so there is nothing more to download.

• Upscale the result. Redraws the result bigger in tiles with my Restore LoRA, to the Megapixels you set. It opens at 8, which is twice each side of a 2 megapixel picture.

All three are off as the file opens.

What you need

• AusBoss nodes 2.10.0 or newer. In ComfyUI-Manager, search "AusBoss". Already have them? Update first.

• ComfyUI 0.38 or newer.

• The models:

  • qwen_image_2.1_turbo_int8_convrot.safetensors → models/diffusion_models · 7.26 GB

  • qwen3vl_8b_int8_convrot.safetensors → models/text_encoders · 9.35 GB

  • qwen_image_2.1_vae_bf16.safetensors → models/vae · 676 MB

• Only for Edit a picture, my Consistency LoRA: qwen-image-2.1-consistency.safetensors → models/loras · 159 MB

• Only for Upscale the result, my Restore LoRA: qwen-image-2.1-restore.safetensors → models/loras · 159 MB

With the switches off it runs on the three model files alone. Auto prompt needs nothing more.

Tips

• The 8 steps are built in. The workflow uses the step list that ships with the checkpoint. A plain KSampler set to 8 steps came out gritty in my tests, so there is no Steps box.

• Do not add a speed LoRA. Viggle Turbo on top of this checkpoint broke the picture: leathery skin and grit.

• With Auto prompt on, type what matters to you. "a record shop at night with a neon sign that says OPEN LATE" is enough. For an edit, type the change: "make it snow".

• Without Auto prompt, for an edit say what changes and what stays: "Change the black t-shirt to a red sweater. Keep everything else the same."

• To change how Auto prompt writes, open the box Write the prompt. Both instructions are plain text.

• The edit size is on the picture node. Set Megapixels in its Resize row. 1 is faster, 2 holds more detail.

• The Consistency LoRA is for edits only. Without it 7 of my 8 test edits sat 2.6 to 10 pixels off the picture. With it all 8 sat within 1.6. Left on for a new picture it makes the result softer, so the switch turns it off.

• If the run stops with "LoRA Loader ... rejected the value for loras", a LoRA file is missing. Download it from the list above.

• A new picture works from 1 to 8 megapixels. On an RTX 5090, 2 megapixels takes about 4 seconds, 4 about 11 and 8 about 34. 8 needs about 25 GB of VRAM.

• For a bigger new picture, raise Latent Size. The Upscale switch is for a picture you already made and want bigger, for an edit, and for a card with less memory: its tiles are 2 megapixels each.

• Times on an RTX 5090 at 2 megapixels. A new picture: about 4 seconds, about 5 with Auto prompt. An edit: about 9 seconds, about 10 to 11 with Auto prompt. The upscale to 8 megapixels adds about 38. The regular model at 25 steps takes about 10 seconds for a new picture.

What it does not do

• An edit keeps the shape of your picture. Ask for a painting and you get your photo with a painted look, not a new drawing.

• Auto prompt fills in what you leave out. It picks the clothes, the light and the framing. Type more to decide more, or switch it off.

• The upscale adds less detail than my Tiled Upscale page. That page runs the regular model and draws more texture. This one is the quick pass.

• The upscale is a redraw. Fine detail is made up to fit, and small lettering can come back misspelled.

• Everything was measured on one RTX 5090.

Credits

Qwen Image 2.1 Turbo is by Alibaba Qwen, under the Qwen Research License: non-commercial use only. The ComfyUI files are from Comfy-Org.

v1.1 (2026-10-12): Upscale the result drew fine dark lines on skin. The upscale now starts from your picture and uses the Turbo's own steps, so skin comes out clean and faces stay the same. If you upscaled with v1.0, run those pictures again. v1.0 (2026-10-09): first release.

Questions or bugs: GitHub issues or the comments here. More from me on X @Zanzibased and GitHub.