Qwen Image Edit 2511 (Free & Open Source) Crushes Qwen Image Edit 2509 & Challenges Nano Banana Pro

[FurkanGozukara]

Qwen Image Edit 2511 (Free & Open Source) Crushes Qwen Image Edit 2509 & Challenges Nano Banana Pro

Full tutorial: https://www.youtube.com/watch?v=YfuQuOk2sB0

Qwen Image Edit 2511 model just published and it is literally competing against Nano Banana Pro at image editing tasks. With native whopping 2560x2560 pixels image output capability and with only 12 steps it is next level. With our installers and specially made Quant FP8 Scaled model, you can run this amazing beast even as low as 6 GB GPUs. In this tutorial, I have compared Qwen Image Edit 2511 with previous successor model Qwen Image 2509 with 12 different unique and hard prompts and cases. Everything is step by step explained and provided.

📂 Resources & Links:

SwarmUI Installer, Model Auto Downloader and Presets: [ https://www.patreon.com/posts/SwarmUI-Install-Download-Models-Presets-114517862 ]

SwarmUI & ComfyUI Setup Guide for Windows: [ https://youtu.be/c3gEoAyL2IE ]

ComfyUI Installer: [ https://www.patreon.com/posts/ComfyUI-Installers-105023709 ]

Image and Video Comparison Slider App: [ https://www.patreon.com/posts/Compare-Images-Videos-Slider-App-133935178 ]

Wan 2.2 Model Training Tutorial: [ https://youtu.be/ocEkhAsPOs4 ]

SwarmUI & ComfyUI Setup Guide for RunPod & Massed Compute: [ https://youtu.be/bBxgtVD3ek4 ]

Upload / Download Big Files Guide for RunPod & Massed Compute: [ https://youtu.be/X5WVZ0NMaTg ]

⏱️ Video Chapters:

00:00:00 Intro to Qwen Image Edit 2511: Replacing a Hat with a Sports Cap Demo

00:00:47 Tutorial Roadmap: SwarmUI, One-Click Installer, FP8 Model, and Hardware Requirements

00:01:43 Case 1 Comparison: Photo Restoration and Colorization (2509 vs 2511)

00:02:36 Resource Info: How to Access the Zip File with All Case Prompts and Images

00:03:05 Case 2 Comparison: Adding a Beard While Maintaining Facial Identity and Realism

00:03:56 Case 3 Comparison: Converting a Real Person into Retro Style Pixel Art

00:04:24 Case 4 Comparison: Combining Three Different Anime Heroes into One Epic Scene

00:05:13 Case 5 Comparison: Removing an Object (Dog) While Preserving Background Details

00:06:00 Case 6 Comparison: Transforming Day Lighting to Realistic Night Time

00:06:38 Case 7 Comparison: Changing Specific Object Color (Rabbit) to Purple

00:07:14 Case 8 Comparison: Changing Clothing Styles While Keeping Body Proportions

00:07:54 Case 9 Comparison: Inpainting a Specific Object (Perfume Bottle) into a Hand

00:08:52 Case 10 Comparison: Zero-Shot Style Transfer (Redrawing a Car Sketch)

00:09:49 Case 11 Comparison: Converting Anime Characters into Realistic Human Renders

00:10:55 Case 12 Comparison: Turning Real Photos into High-Quality Anime Style

00:11:32 Installation Step 1: Updating ComfyUI to Support New Custom Nodes

00:12:45 Installation Step 2: Setting up the SwarmUI Model Downloader Script

00:13:08 Installation Step 3: Downloading the Qwen Image Core Bundle and Verification

00:13:35 Installation Step 4: Updating SwarmUI to the Latest Version via Batch File

00:13:52 Installation Step 5: Importing the "Amazing SwarmUI Preset Version 40"

00:14:47 Configuration: Setting Base Resolution to 1920x1920 for Best Performance

00:15:33 Live Workflow 1: Single Image Editing and "Closest Aspect Ratio" Strategy

00:16:32 Hardware Monitoring: RAM Usage, SageAttention, and 6GB GPU Optimization

00:17:36 Live Workflow 2: Combining Multiple Input Images with 16:9 Aspect Ratio

00:18:35 Performance Analysis: Generation Speed (Steps/Sec) and Power Usage on RTX 5090

00:19:44 Live Workflow 3: Fast Single Image Edit and Reference to Wan 2.2 Tutorial

00:20:41 Tools: Installing and Using the Image and Video Slider Application

00:21:56 High-Res Experiment: Pushing Resolution to 2560x2560 for Maximum Quality

00:22:28 High-Res Monitoring: Analyzing VRAM Usage and Generation Time Increase

00:23:18 Final Verdict: Why 2560x2560 is the Best Resolution for Qwen Image Edit 2511

In this comprehensive tutorial, I demonstrate the incredible power of the new Qwen Image Edit 2511 model, comparing it directly against the previous 2509 version across 12 distinct test cases. You will see how this model achieves "Google Nano Banana Pro" levels of quality in photo restoration, style transfer, object replacement, and multi-image composition.

I have prepared a custom FP8 scaled version of the model that allows it to run on GPUs with as little as 6GB VRAM while maintaining near-BF16 quality. This guide covers everything from a full 12-case comparison to a complete step-by-step installation guide using SwarmUI and ComfyUI.

📂 In this video, you will learn:

How to install Qwen Image Edit 2511 with a one-click installer.

How to use the new "Amazing SwarmUI Preset v40" for optimal results.

A side-by-side comparison of 2509 vs. 2511 (Restoration, Pixel Art, Anime-to-Real, and more).

How to combine up to 3 images into a single generated scene.

Optimal resolution settings (1920x1920 vs 2560x2560) for maximum accuracy.

How to use the Image and Video Slider Application for professional comparisons.

Video Transcription

• 00:00:00 Greetings everyone. What you are seeing  here right now is Qwen Image Edit Model  

• 00:00:05 2511. This is original input image, prompt  is replace hat with a sports cap and this is  

• 00:00:11 output. I mean look at these details.  Look how it achieves perfect quality,  

• 00:00:17 perfect accuracy. This is Google Nano  Banana Pro level. You see it is doing  

• 00:00:22 the task with keeping image as it is  with amazing quality, amazing accuracy,  

• 00:00:28 amazing high resolution. So today I will be  showing you how to use Qwen Image Edit 2511  

• 00:00:35 model and I will make a detailed comparison with  the previous Qwen Image Edit 2509 versus Qwen  

• 00:00:42 Image Edit 2511. You will love this tutorial,  watch to the very end to learn everything.

• 00:00:47 Greetings everyone. Today I am going to show  you how to use Qwen Image Edit 2511 model with  

• 00:00:54 most easiest and most accurate way on your PC  or on cloud by using the SwarmUI. Moreover,  

• 00:01:02 I will compare it with the Qwen Image Edit  2509 model so you will see how much they  

• 00:01:09 have improved. With using our Qwen Image  Edit 2511 preset that is only 12 steps,  

• 00:01:17 you will be able to generate amazing quality  images and with our one click downloader you will  

• 00:01:23 download the accurate models right away. With our  one click installer you will be able to download,  

• 00:01:29 install, setup and everything will work as low  as 6 gigabyte GPUs with amazing quality. Why?  

• 00:01:37 Because I have prepared an amazing quality  FP8 scaled version of the model for you.

• 00:01:43 So let's begin with the comparison of the Qwen  Image Edit 2509 to 2511. As you can see this  

• 00:01:51 is a restoration, photo restoration. Let me  open the original and let's see the 2509. So  

• 00:01:58 this is the original image. 2509 epically fails  as you are seeing right now. I have generated 4  

• 00:02:05 seeds of every image and I have picked the best  one for every model. Let's open the 2511 and as  

• 00:02:12 you can see 2511 is able to colorize and restore  photo. So let's see comparison. This is the 2511,  

• 00:02:23 you can see the name here, restored photo, this  is the original photo. So the 2511 handles this  

• 00:02:30 task amazingly. Moreover, I have shared  all the used images, prompts in this zip  

• 00:02:36 file. You see it is here when you scroll down  a little bit you will see it. Inside this zip  

• 00:02:41 file you will see the case prompts. Let me open  it. So these are the case prompts. Case 1 prompt  

• 00:02:47 and when you enter inside the cases you will  see all the images. This is the base image,  

• 00:02:52 this is the 2509 result and this is the  2511 result. So for 2509 I have used our  

• 00:03:00 previous best preset. Now it is removed  our presets because 2511 is the best one.

• 00:03:05 So let's continue with the next case, image 2.  And this is the result of the 2509. So this is  

• 00:03:12 our input image and this is the generated image.  As you can imagine the prompt is adding beard.  

• 00:03:19 Let's see the 2511. So the 2511 is keeping the  face as it is as you can see. It is perfect. And  

• 00:03:27 it is adding the prompt that we wanted. There is  a massive difference between 2509 to 2511. 2511 is  

• 00:03:36 next level. One another thing I have made is that  I did set the base resolution of the model to 1920  

• 00:03:44 to 1920 for these experiments. It is working  amazing even at this resolution. Don't worry,  

• 00:03:50 I will show everything how to install, download  and setup after we completed the comparison part.

• 00:03:56 So the image 3 is this one. This is our base  image and let's see the 2509. The prompt is  

• 00:04:03 turning the person, the image into pixel art.  As you can see 2509 epically fails compared  

• 00:04:10 to 2511. 2511 is perfectly converting image  into retro style pixel art as you can see.  

• 00:04:18 This is really really good. This is Nano  Banana Pro level. The 4th experiment is  

• 00:04:24 combining 3 images. Which 3 images? So I have  provided all 3 images, all of these images,  

• 00:04:32 and I asked AI to combine them. How I asked?  I said that put all 3 anime heroes into same  

• 00:04:40 epic anime scene. So let's see the generated  results. The left one is 2511. As you can see  

• 00:04:47 it did put all 3 heroes into the scene perfectly  accurately as you can see. However 2509 epically  

• 00:04:56 fails. It puts 5 characters, the characters  are changed. I mean this character doesn't  

• 00:05:00 exist. This is not the Saitama. This is Naruto  but somewhat different. This is a little bit  

• 00:05:06 more accurate. So as you can see there is  a massive difference between 2511 to 2509.

• 00:05:13 Let's move to case 5. So the case 5 is removing  the dog from the scene. Let's see the result  

• 00:05:19 of 2509. So the 2509 changes the image almost  entirely as you can see. The dog is removed but  

• 00:05:27 the quality is very low. And let's see the result  of the 2511. So the 2511 keeps the image as it is  

• 00:05:35 as you can see with minimal amount of changes and  it removes the dog perfectly fine. You are seeing  

• 00:05:41 some quality loss because the original image  resolution is very high. It is like 20 megapixel.  

• 00:05:47 Therefore it is very high resolution. However our  outputs are like 4 megapixel. Therefore we lose  

• 00:05:53 quality at very high resolution but this is really  good. I mean this is like Nano Banana Pro level.

• 00:06:00 Let's see the case 6. So the case 6 is turning  this image into night time. Let's see the result  

• 00:06:07 of 2509. So this is 2509. I mean this looks so  artificial. It changed the image aspect ratio,  

• 00:06:17 zoom. So this is 2509. Let's see the 2511. The  2511 handles it much better. I mean there is  

• 00:06:25 a magnificent difference between 2511 to  2509. So you see it keeps everything as  

• 00:06:33 it is. It is almost perfect. So the case  7 is turning rabbit into purple color. So  

• 00:06:38 this is the original image and this is the  2509. You can see that the image changed,  

• 00:06:44 now it is not anymore the real photo,  everything changed. But when we see the 2511,  

• 00:06:51 it is next level. I mean it is trying to keep the  image as it is with minimal amount of changes. Of  

• 00:06:57 course the resolution is dropped because the  input image is high resolution. However it  

• 00:07:02 handles the task. And you are seeing it in 4K  resolution. So when you see it on your mobile  

• 00:07:08 devices or tablets or some other places you  wouldn't even notice that this is AI made.

• 00:07:14 Case 8 is changing the clothing of the  person. So this is 2509. It changed the  

• 00:07:20 clothing. However the realism reduced, the  image proportions reduced. So this is 2509.  

• 00:07:26 Let's see the 2511. So the 2511 is much  more realistic. It is keeping the details,  

• 00:07:32 the accuracy is amazing. I prefer the 2511 all  the way. You can see that it is keeping the image,  

• 00:07:40 making minimal amount of changes. Still  I can see that the resolution is dropped  

• 00:07:44 because the original image is very high  resolution. However still this is amazing.  

• 00:07:49 This is excellent level. This is running  on my PC with as low as 6 gigabyte GPUs.

• 00:07:54 So the case 9 is making person holding a perfume  bottle. So I am going to make the person to hold  

• 00:08:02 this bottle. Let's see the result. So this is  2509. It is one of the best where 2509 did. It is  

• 00:08:10 pretty good. But nothing sort of the 2511. Let's  see the 2511. So this is the 2511. I see that the  

• 00:08:19 colors are not perfect in this one. However with  more generations probably I can get better colors.  

• 00:08:24 But the accuracy, the face realism, the face  shape, it is much better than the 2509. I mean  

• 00:08:31 let's make a comparison. So 2509, 2511. Let's  see base image. So this is how 2511 is making  

• 00:08:41 this image holding this bottle. Very good in my  opinion. And on low resolution you wouldn't even  

• 00:08:47 notice the difference. At very high resolution it  is still pretty good. This is a very hard task.

• 00:08:52 The test case 10 is very hard. I want the model  to redraw this input car with the style of this  

• 00:08:59 input image. So let's see the result of the  2509. The 2509 epically fails. I mean it puts  

• 00:09:06 the car into the image. It doesn't redraw  it with the desired style. So basically we  

• 00:09:12 are doing a zero shot style transfer. Let's see  the result of the 2511. The 2511 redraws image  

• 00:09:22 with perfect accuracy with the desired  style. I mean this is next level. This  

• 00:09:28 is really good. However it put the background  but it is perfect. I mean the car, the shape,  

• 00:09:33 the style, everything is so perfect. This model  shines. With the background removers you can  

• 00:09:38 also remove the background if you want. Perhaps  with prompting the background can be removed as  

• 00:09:43 well. But this is perfect. This model is very  very strong. This is Nano Banana Pro level.

• 00:09:49 So the case 11 is one of the hard tasks. Turning  image into a real photo. This is an anime image,  

• 00:09:56 a famous character and I want it  to turn it into a real image. So  

• 00:10:00 this is 2509. This is what it makes best. I  tried several prompts and I am always fair,  

• 00:10:07 I tried same on the both models. So this is  the result of 2509 from this image into this  

• 00:10:14 realistic image. Let's see the 2511. The 2511  does much better job. This is like a realistic  

• 00:10:21 render. Yes it is not still a real person.  However it is really really good compared  

• 00:10:27 to 2509. The 2509 is much lesser detailed, the  accuracy is lower. However 2511 is much better.  

• 00:10:37 You can play this character in a movie, movie  scene. Perhaps some with additional upscales,  

• 00:10:43 maybe image to images it can be made certainly  much better realistic. But the output is really  

• 00:10:49 good when you consider the base image from this  image into this image. It is really really good.

• 00:10:55 Okay case 12 is turning real image into anime  image. So 2509 epically fails. This is not like  

• 00:11:04 an anime image. You can see that it fails to  zero shot stylized image generation. However  

• 00:11:12 let's see the 2511. The 2511 is perfect. As you  can see it turned image into anime with zero shot,  

• 00:11:20 with perfect details, perfect accuracy.  Everything is so good. So the 2511 is  

• 00:11:26 much bigger update than people realize. It is  really really powerful. It is really amazing.

• 00:11:32 So how we are going to use this model like I do?  As usual I have everything in this amazing post.  

• 00:11:39 If you don't know how to install SwarmUI,  ComfyUI and how to use, please follow this  

• 00:11:44 tutorial. First of all we will update our ComfyUI  installation. The ComfyUI installation link will  

• 00:11:50 be also in the description of the video and in  the pinned comment, it is also here. So let's  

• 00:11:55 download the latest ComfyUI zip file. This is  mandatory. Why? Because this Qwen Image Edit  

• 00:12:01 2511 quant scaled FP8 requires a special custom  node and my installer will auto install it. This  

• 00:12:09 is not natively supported by the ComfyUI yet but  hopefully it will be. So as usual move the latest  

• 00:12:15 zip file into your installation. You can make a  fresh installation. We have upgraded installers  

• 00:12:20 to the UV so it is much faster. Extract and  overwrite all files. This is important. Then all  

• 00:12:26 you need to do is just run the Windows install  or update ComfyUI. It will quickly update the  

• 00:12:32 ComfyUI to the latest version. It will install  the new necessary custom nodes. Then we will be  

• 00:12:39 able to use the latest model in our SwarmUI. So  everything is updated, it took like 30 seconds.

• 00:12:45 Then as a next step download the latest  SwarmUI model downloader zip file. It is  

• 00:12:51 11.4 at the moment. Then move it into your SwarmUI  installation folder as usual. You can also make  

• 00:12:58 a fresh installation. Extract and overwrite  all files. This is super important. You need  

• 00:13:04 to overwrite all files. First of all you need  to download new model. So for downloading new  

• 00:13:09 model Windows start download models app.bat  file. Our downloader is extremely fast,  

• 00:13:14 extremely robust. It verifies all of the  downloaded files. We have upgraded the  

• 00:13:19 interface significantly. So I recommend you  to download Qwen Image Core Bundle to not  

• 00:13:24 miss any of the necessary models. You see  it is only 56 gigabytes. Moreover it will  

• 00:13:29 skip already downloaded files. After that  you need to run the Windows update SwarmUI  

• 00:13:36 to get the latest SwarmUI version. This  is also pretty fast and it will start.

• 00:13:41 After that you need to update your  presets. You can use import preset,  

• 00:13:46 choose file and from there pick the latest  preset, Amazing SwarmUI Preset version 40.  

• 00:13:52 Overwrite and import. Alternatively you  can also use our preset importer which is,  

• 00:14:00 when your SwarmUI is running you click the Windows  preset delete import. It will backup your preset  

• 00:14:05 and it will import, click yes and all imported.  Then click refresh. This will do a clean preset  

• 00:14:12 import. It will delete all previous ones and it  will import the newest ones. I recommend this one.  

• 00:14:18 Then you will have the newest Qwen Image Edit  preset here. The downloader will download  

• 00:14:23 them with 16 connections and as soon as files  downloaded it will merge the parts and verify  

• 00:14:30 hashes as you are seeing. Wait for all downloads  to be finished and once the all downloads have  

• 00:14:36 been completed go to models click refresh. Go  to Loras click refresh or restart your SwarmUI.

• 00:14:42 Then I recommend you to change the resolution.  Because higher resolution works better with this  

• 00:14:47 model so click this hamburger menu, click edit  metadata. Set this as 1920 to 1920. Maybe even  

• 00:14:55 higher may work. So after that it will auto set  the resolution and aspect ratios accordingly.  

• 00:15:01 Then as usual we click quick tools, reset params  to default, refresh. Then let's go to presets.  

• 00:15:08 Find the Qwen Image Edit new preset. You can  search from here Qwen and you will see that  

• 00:15:14 Qwen Image Edit 2511 direct apply. You should see  that it is selected this Qwen Image Edit quant  

• 00:15:22 model and this Qwen Image lightning Lora. This  is the best working one don't worry about that.

• 00:15:28 So how you are going to use the model? Let's  demonstrate with 2 examples. Click this plus  

• 00:15:33 icon. We are going to upload prompt image.  So from the case images let's do this case.  

• 00:15:38 So this is our input image. To set the accurate  resolution I recommend you to use this strategy.  

• 00:15:44 So from init image choose file, choose the  same image, click res use closest aspect  

• 00:15:51 ratio. You can also use exact aspect ratio or  you can use exact resolution. However you need  

• 00:15:57 to test. You see the exact resolution is very big  so I am going to use closest aspect ratio. If you  

• 00:16:01 make images exactly this resolution yourself  manually it can work even better. So try to  

• 00:16:07 crop images yourself to this resolution and use  that way. Then disable this. We don't use init  

• 00:16:13 image. This is only for setting the resolution.  Don't forget that. After that type your prompt,  

• 00:16:19 something like this. Then generate as many  as images you want. Let's generate 4 images.

• 00:16:24 Let's also see nvitop. So nvitop will show us  what is our GPU usage. I have RTX 5090. So the  

• 00:16:32 generation already starting. I recommend you  to not run any other application while SwarmUI  

• 00:16:38 is running with ComfyUI because recent ComfyUI  versions have problems for RAM management. It  

• 00:16:45 cannot properly detect your used RAM memory.  Therefore I recommend you to not run any other  

• 00:16:51 applications. So it is using 50 percentage  of my RAM memory right now. It is using my  

• 00:16:56 23 gigabyte. But don't worry since we are  using ComfyUI backend as I have shown in  

• 00:17:01 other tutorials with SageAttention and latest  version it is properly doing block swapping,  

• 00:17:07 CPU offloading. So with 6 gigabyte  GPUs you can run this model.

• 00:17:11 Okay the first image has been generated. It is  amazing quality. I mean look at this quality.  

• 00:17:17 This is really good. The resolution is really  high resolution. You see it is 2288 to 1568.  

• 00:17:25 We may even go with higher resolution depending  on your input image but the result is amazing.  

• 00:17:31 What about combining multiple images? Let's clear  image. Click this plus icon. So you need to upload  

• 00:17:37 every image. So let's upload our these 3 images.  Second one and the third one. As far as I know the  

• 00:17:44 model supports up to 3 images. I didn't try 4 but  you can try it. So for this one I am going to use  

• 00:17:51 this prompt that we have. Okay and generate. Oh  I have forgotten to change the aspect ratio. So  

• 00:17:59 you need to set your aspect ratio according to  your desired image. For this one since this is  

• 00:18:04 a combination of multiple images I am going to  use 16 9. So this will be the resolution. You  

• 00:18:10 probably can combine these 3 images with other  resolutions, other aspect ratios as well. But  

• 00:18:16 this is what I want. You can choose anyway. We  are only doing 12 steps therefore the images  

• 00:18:21 are really fast. I mean let's see the generation  speed. It should be done under 60 seconds on my  

• 00:18:28 GPU at this massive resolution. Okay it is getting  done right now. Let's see the step speed from the  

• 00:18:35 debug. So it is like 8.5 second per IT. You can  reduce the steps count. It is totally fine but I  

• 00:18:44 prefer 12. You can go as low as 4 steps because  we are actually using 4 steps Lora but 12 steps  

• 00:18:51 is yielding the best quality. Moreover when you  use the base resolution not this bigger resolution  

• 00:18:57 it will be even faster but we are currently doing  4 megapixel output. Therefore it is taking more  

• 00:19:03 time than what it should take normally. However  the quality is amazing let's see. You can also  

• 00:19:09 monitor your power usage. You should be using up  to maximum power available like 575 to 575. This  

• 00:19:18 means that it is working with 100 percentage  performance. There is no performance loss.

• 00:19:23 Okay and the image has been generated. This  image took longer than the other ones. Why?  

• 00:19:29 Because as you input more images it takes more  time. Therefore this image took 111 seconds.  

• 00:19:37 This is why this one took more time. If I just  use a single image, let's make another demo. For  

• 00:19:44 example this image. I will say that replace  hat with a sports cap. Let's see the result.  

• 00:19:52 This will be much faster than using 3 images.  When you use more image as an input it will  

• 00:19:58 take more time. But this is the logic of how  you use multiple images. You see it is getting  

• 00:20:03 generated much faster. It is 3.5 second IT right  now. When it was 3 images it was 8.6 second IT.  

• 00:20:12 So you see it is faster than twice. So this was  our input image. Let me show you. By the way this  

• 00:20:18 image has been generated with our recent Wan 2.2  tutorial. It is an amazing tutorial let me show  

• 00:20:24 you. So I recommend you to watch our latest Wan  2.2 tutorial on our channel. This is an amazing  

• 00:20:30 tutorial. And the generated image is here.  I mean look at this. It is almost perfect.

• 00:20:36 So how I am comparing these images? Let's download  this image. For image comparison I have developed  

• 00:20:42 an application. It is called as Image and Video  Slider Application. Go to this link. It will be  

• 00:20:47 also in the description of the video. Download  the latest zip file. Move the file wherever you  

• 00:20:52 want to install. For example let's install  into our tutorial folder. Okay extract here.  

• 00:20:58 Then install update application. It will just  install, it is very fast. Then run the start  

• 00:21:04 image comparison. This is a very advanced  application. You can drag and drop multiple  

• 00:21:09 files for comparison like all of these files.  I can drag and drop them or select from here  

• 00:21:14 and it will sort them by name so it will be very  easy. You can also change them from here. 1 PNG,  

• 00:21:20 this one 2509. You can also look from here  or you can full screen like this and you  

• 00:21:26 can compare. This is an amazing application.  So let's compare our latest generation. So I  

• 00:21:31 am selecting both of the files. Let's do  a full screen. Okay the right one is the  

• 00:21:36 original one. The left one is the new Qwen Image  Edit 2511. I mean this is really good quality.  

• 00:21:45 It minimally changes the image. The original  image is very high resolution that is why you  

• 00:21:49 see it is more crisp and sharp but this output is  really really good compared to previous models.

• 00:21:56 I wonder something, if I make this even higher  resolution what happens? So I am going to change  

• 00:22:01 the resolution to 2560 to 2560. Then select  another model and select model so it will  

• 00:22:08 update the input resolution. This is same as the  original image resolution. Let's see the result.  

• 00:22:15 So obviously higher resolution will take more time  but it should use the same VRAM let's see. Yes  

• 00:22:22 it is using the same VRAM amount. It will just  take more time. I expect quality to be higher.  

• 00:22:28 Let's see. Now it is taking 7.35 second. It was  3.53 second because we increased the resolution  

• 00:22:36 significantly. And you see my model is fully  working with SageAttention. Both on RTX 3090  

• 00:22:43 Ti and RTX 5090. It is working on all GPUs because  I have compiled this SageAttention myself. I have  

• 00:22:51 generated this FP8 scaled myself. This is really  high quality almost BF16 quality. Wow it looks  

• 00:22:57 like it is going to work. Let's see. The other  image was generated in 45 seconds. It was still  

• 00:23:04 big resolution 4 megapixels. Okay the new image  is getting done. All VRAM management is being  

• 00:23:10 made by the ComfyUI. Okay it is done. Wow. Wow  amazing. Let's download this and let's make a  

• 00:23:18 comparison. Wow look at this image. You see when  you use higher resolution it is almost perfect.  

• 00:23:24 From this image to this image. I mean there is  almost no loss of quality. It did the job and  

• 00:23:30 it did amazing. I mean look at this quality  from this to this. There is almost no quality  

• 00:23:36 loss. It minimally changed the image. It is still  realistic as it is. So this is an amazing quality.  

• 00:23:43 Therefore I recommend you to set the resolution  to 2560 to 2560 and it will be amazing quality.

• 00:23:50 So this is it. I hope you have  enjoyed. Please like, subscribe,  

• 00:23:54 leave a comment and ask me any questions  that you have. So it appears that the best  

• 00:24:00 resolution for this model is 2560  to 2560 in high resolution images  

• 00:24:06 to keep the maximum quality, maximum  accuracy. See you later hopefully.