Qwen Edit 2511: max quality outputs, minimal workflows + lots of info — Qwedit 2511

Author notes from Qwedit 2511 - 2 image.json · selected release

## Main Model Goes in models/diffusion_models FP8 recommended, it's much higher quality than the Q8 GGUF. [qwen_image_edit_2511_fp8mixed.safetensors](https://huggingface.co/Comfy-Org/Qwen-Image-Edit_ComfyUI/resolve/main/split_files/diffusion_models/qwen_image_edit_2511_fp8mixed.safetensors) Requires 22GB VRAM --OR-- If you're RAM limited [GGUF Versions](https://huggingface.co/unsloth/Qwen-Image-Edit-2511-GGUF/tree/main) Requires [ComfyUI-GGUF](https://github.com/city96/ComfyUI-GGUF) extension, load the model with "UNET Loader" node ## Text Encoder [qwen_2.5_vl_7b_fp8_scaled.safetensors](https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/resolve/main/split_files/text_encoders/qwen_2.5_vl_7b_fp8_scaled.safetensors) Goes in models/text_encoders ## VAE [qwen_image_vae.safetensors](https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/resolve/main/split_files/vae/qwen_image_vae.safetensors) Goes in models/vae   ## Non-direct Model Repo Links https://huggingface.co/Comfy-Org/Qwen-Image-Edit_ComfyUI/tree/main https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/tree/main

Author notes from Qwedit 2511 - 2 image.json · selected release

Always use 2.5 CFG and 3.0 Shift. Do 30 steps for maximum sharpness, but I recommend only doing 20 steps then downscaling->upscaling after to restore sharpness. Saves time, and deals with the Qwen halftone pattern. You can also go as low as 10 steps, which makes it run quicker but lowers the quality; good for quick testing.

Author notes from Qwedit 2511 - 2 image.json · selected release

This confusing little set of reference latents is how we feed the input images into the model. It only looks confusing because of the switch earlier that lets you turn off the double input.

Author notes from Qwedit 2511 - 2 image.json · selected release

Note: read the general prompt tips in the single-image workflow for the basics. This box only covers multi-image prompting. Most of the time you'll need to describe your images briefly before your actual prompt. This is because this workflow skips the usual method Qwedit uses to 'figure out' which image is which - so we have to tell it. You MUST phrase it as "Picture 1: ", "Picture 2: ", etc. Afterwards you refer to them as "picture 1" and so on when asking for the edits. Prompt example 1 - > Picture 1: a man wearing a professional suit. Picture 2: a man wearing a hawaiian shirt and jeans. > > Make the man in picture 1 wear the outfit from picture 2. Prompt example 2 - > Picture 1: an empty rooftop bar. Picture 2: a close up of a woman in a restaurant. > > Make a new shot of the woman from picture 2 sitting on a chair at the rooftop bar from picture 1.

Author notes from Qwedit 2511 - 2 image.json · selected release

Note: image 1 is on the right, image 2 is on the left. **Enhance with Double Ref:** > When true, the input images are given to the model **twice**. *On average*, this noticeably increases the quality of the results in every way. Better prompt adherence, better detail, better consistency. > Warning: generation takes WAY longer with this on. Multi-ref inputs for edit models already take extra long, and the double-ref method amplifies this 2x. The trade-off for speed vs quality is very high on this one. **Scale Input Images:** > When true, the input images are scaled by their **longest edge** to the lengths specified in the corresponding input boxes. > Additionally, all of their dimensions are forced to be divisible by 16 (required for Qwedit) regardless of whether scaling is set to true or false. It will **crop** the images to accomplish this, but you can change it to pad the images with black instead by changing the "fit" in the LayerUtility nodes above the images. > It's a good idea to stick with common values, such as: 1920, 1600, 1440, 1280. > Qwedit can *easily* handle images up to 2.5 megapixels (e.g. 1920x1440), but higher than that it can be iffy. Also note that a 2.5 mpx image takes 2.5 times longer to process than a 1 mpx image. > Be warned, the slowdown gets crazy if you input multiple HD images and use the double ref method at the same time. Like 10 minutes to process three 1920x1440 inputs on an RTX 5090. **Custom Output Dimensions:** > When true, uses the custom dimensions you specify in the corresponding box for the output image. > When false, it uses the same dimensions as the first input image (after it's been resized). > It helps to use common aspect ratios like 2:3, 3:4, 4:5, 16:9, etc. > The size of your input image still matters. If your input is 1024x1024 and your output is 1440x1440, details that carry over from the first image may be blurry or lack detail. Here's a handy resolution chart for common aspect ratios, all values divisible by 16: [https://ibb.co/S7rKtzfT](https://ibb.co/S7rKtzfT)