Insane Local Testing Qwen-Images-2.1 on my 8GB GPU

Author: Prompt Engineer 48 · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Insane Local Testing Qwen-Images-2.1 on my 8GB GPU · Channel: Prompt Engineer 48 · 16:38 · Watch on YouTube

Topics: AI developer tools, AI content generation, Local and open-source models

Summary

The video demonstrates running Qwen Image 2.1 locally on an RTX 4060 laptop with 8GB of GPU memory, using a 4-bit GGUF quantized model in ComfyUI. The creator walks through setting up ComfyUI, downloading the model files and workflows, and generating images from text prompts. Examples include posters, infographics, product photography, transparent assets, and lettering. The video also shows image editing with reference images, including changing a woman’s white dress to a supplied red dress.

Key points

The demonstration uses an RTX 4060 with 8GB of GPU memory and a 4-bit GGUF quantized version of Qwen Image 2.1.
The creator uses ComfyUI Desktop to set up the local workflow.
The required model components include a diffusion model, text encoder, and VAE; the video also provides text-to-image and image-edit workflow JSON files.
ComfyUI can download missing model files directly when they are detected in a workflow.
The text-to-image workflow includes prompt enhancement, text encoding, sampling, and VAE decoding.
Generated examples include a concert poster, an infographic, a product photo, a red panda, transparent assets, and exact text.
ComfyUI workflows can be exported as API JSON for batch generation.
The image-edit workflow accepts reference images and is used to replace a woman’s white dress with a red dress.
The dress-editing generation takes about 126 seconds, roughly twice as long as text-to-image generation in the creator’s test.

Chapters

Start	Chapter	Summary
00:00	Qwen Image 2.1 and local testing	The creator introduces Qwen Image 2.1, its capabilities, and the plan to test it on an 8GB RTX 4060.
01:02	Choosing the model and setting up ComfyUI	The creator selects an uncensored GGUF model and starts a ComfyUI Desktop instance for local use.
03:05	Downloading workflows and model files	The video covers the required quantized diffusion model, text encoder, VAE, and text-to-image and image-edit workflow files.
08:20	Understanding the workflow and generating images	The creator explains the text-to-image workflow and runs prompts to generate several sample images.
12:30	API export and image editing setup	The video shows how to export a workflow as API JSON and how to load the image-edit workflow and its missing models.
14:01	Testing image editing and final results	The creator prepares reference images, changes a woman’s white dress to red, and shares the generation time and result.

Quotes

everything here is running on my local system, which is an RTX 4060 GPU — 00:00

it can change the subject’s expression when retaining the transparent body. — 04:40

if you’re trying to do batch generation, this is the way out. — 12:30

This took about 126 seconds, which is almost double the time — 16:07

Concepts and entities

Qwen Image 2.1, ComfyUI, ComfyUI Desktop, GGUF, RTX 4060, API do ComfyUI, ChatGPT

Full transcript (original language)

00:00 All right, in this video, we will see Kwen Images 2.1 and you can see that it is able to change the dressing here. There are other experiments that we’ll do here as well. You can see the images generation that we have done. Now, mind you that everything here is running on my local system, which is an RTX 4060 GPU and let’s go ahead and test this out. They have opened Weight Image Kwen Image 2.1. It’s an image model in the Kwen family that balances generation quality, inference efficiency, and cost. It has 7 billion

00:31 parameters only in the visual component generation component and there are four key improvements: compact and efficient, native transparency, unified creation and editing, versatile editing, and had realistic textures and refined aesthetics. Really good and you can see that it stands at number seven here and I really like it. This is a lightweight architecture. It’s visual generation component contains 32 single stream diffusion transformers layers and 7 billion parameters and it’s really amazing. So, let’s see these examples,

01:02 but more than the examples, I’m going to go ahead and check it out on my own system. Now, for trying out the models, I went through the model section here and I was able to see Kwen Image 2.1 here, which is a 48.4K downloads, but the one that really interested me was this one, which is an uncensored model and it’s a GGUF file and it has again a good number of downloads here. So, if you go to that file by Evan Zerbs, it’s an uncensored model and there are some GGUF quantized

01:32 versions here as well that will be very beneficial for this 8GB GPU laptop that I have here. Now, for running this, we will use ComfyUI. Now, for setting up ComfyUI, you can go ahead and go through the usual style of doing this on a Linux system, but now ComfyUI have the desktop feature as well, so you can go ComfyUI desktop and you can download this desktop and run it on your own system. That’s a and very easy way to use this vision language models. So, let’s go ahead and download this on desktop. So,

02:04 once you download this, you will have something like this comfy desktop. And let me open that up on my system. So, if there are any updates you need to update, you can sign it up. I’m already signed up uh with my mail. Doesn’t matter if you’re signed up or not. But, what you can do is now go ahead and create a new instance here. And let’s go ahead and name it, for example, let’s say YouTube one. And public builds, go to a public builds here. And you can see that we’re using this Nvidia recommended here starter template. It’s okay. You

02:34 can just go ahead and click on. You can see detected GPU Nvidia here. And then the required memory is 33 GB. The required is 4.96 GB. And we have 33 remaining, which will work really fine. So, I can skip all these. And you can see that it’s unpacking the installation here. So, it’s getting the entire setup ready. And in the meantime, what you can do is I can go ahead and go ahead and get that JSON file. So, if you go here to the uncensored GG UF, you will have this structure. So, these are the files which are required for the models. Under

03:05 the diffusion models, we need the Kuan image 2.1 a UC quantized 4KM. This is chosen as a GG of quantization. And depending on my laptop, this would be the good option. Text encoders, we’re going to use the text encoders the safe double Bfloat 16 or the int 8 as well. Then the VAE, we would need the Kuan image 2.1 VAE BF16 safe tensors. So, I can see this is almost getting ready. And now what you can do is you can get these two workflows. The first is the text to image here. And then we

03:36 have the image edit as well. So, I’m going to go ahead and open it up on a new tab. This is the text to image. I can go ahead and download this. So, I’m going to put it on a new folder here. So, let’s download this. I’m going to download the image edit as well. So, right click open on a new tab here. and then we can download the image edit JSON as well. So, let’s go there to the folder here. So, you can see that we have two JSON files that will help you load the models here. Essentially, you can create your own, but these are really good to have, and you can

04:08 specially speed up your workflow and test it out as soon as possible. So, while this is loading in the meantime, what you can do is I’m going to go to chat GPT and get some prompts here. So, this is the page. I’m going to give the page to chat GPT, and can you give me 20 prompts to test out this, okay? So, it’s going to walk through this, and this will be ready in just a few minutes. It does need time. If you need another GPU, you can go ahead and always go ahead and get cloud GPU, but for this local system, I really would

04:40 like to test this 4-bit quantized model here. So, look at the few examples. So, it has a native transparency, unified creation and editing, and here the examples of a transfer images. It looks really crisp. It can be used for a lot of projects, and you can see by unifying generation and editing, when image 2.1 also supports editing transparent images directly. It can change the subject’s expression when retaining the transparent body. So, it’s from like boring to like excited here. And you can see that in the transparent layer itself, you can change the text from blue to when image. That’s really great

05:11 capability to test. This is another uh the capability extends to real photograph as well. Given an RGB image, the model can extract the desired subject as an RGB layer with transparency, easier to make it reusable in subsequent designs. So, it was able to extract this leaf here, and now you can use this leaf uh somewhere else. We have versatile editing as well. Up to 10 reference images can be put. So, you can see that this person here, this red uh shirt person, and this person uh this lady, and this person in a coat here,

05:43 and this person all on the same frame, which is really good. A group photograph generated from six portrait images. This is another thing that you can go ahead and test it out. So, I guess this is done. We have our system ready here. Now, what I’m going to do is I’m going to take that JSON file that we had. I’m going to just paste it or pull it here. Then, you’ll see that we have these here and we can see that we need to put these models here. Now, if you click on this location, you can see that this is missing this model. So, what I can do is I can just download it here and this

06:14 will download the models here. So, once it is downloaded, what you can do we can just refresh it and you will have all the models here. But, on the left you can see we have this aspect ratio of square square photo 3:2, 3:4, 4:3, then we have 9:16, 16:9, and ultra wide. I’m going to go with square for now and you can change the megapixel as well, but it will need definitely some time. CFG is one and the steps is 25, sampler Euler, scheduler simple, and this is the things

06:45 that we need here. So, while this is downloading, you can see that it’s downloading here about 8GB file. It will definitely take some time. But, if you go back to the workflow here and let me show you here, these are the files that you need. So, under models diffusion models, so if you go and look inside your system for the models, you need the path obviously. So, you go to desktop settings here and then go to storage here and you can see that this is the path here. So, I’m going to open the path and copy the path here. Let’s

07:17 close everything else and then go to my computer or file explorer and then go ahead and go to this models here. So, this is the folder where everything is available. So, inside the models we need diffusion models and inside the diffusion models we need to keep the Q8. So, inside the diffusion models we need to keep this model of UCQ4KM. All right? So, to get this model you can go ahead, go to the top here, and then you can get this model. So, you see a Q4KM. This one is a model. So, just

07:49 right click, or you can just open it on the new tab, and then just go ahead and download this model. Now, for downloading this, you can just click on this download button. This is going to download on your local system. You can just point it to the exact folder that you want here, artificial models, and then you can just download it here. Now, same goes for the other things as well. So, the other things being the text encoders, and you can get it from here, and then the VAE as well. You can get it from here. But, if you don’t want to do this, you can just go ahead and

08:20 there will be errors here. You can go ahead to the errors and just click on download the missing model. It’s going to download the missing model for you, all the models specifically, and then you can just right click here, and then you can just left click, and you can just select the models for getting the generation started. Okay? So, while this is generating, I’m going to go back and see if my prompts are ready. So, I go to ChatGPT, and you can see that we have these different things that we can go ahead and test out. So, we will test it out in

08:50 just a moment. But, if you go back, and you can see this block, this is a text to image quant image 2.1, and if you just click here, you’ll be brought to the more exploded things exploded view that we have here. So, we have the diffusion model here. We are loading the clip, loading the VAE, and then we have the empty latent space generation here. So, the width and height are taken from the inputs here, and then we have the text encode here, and then we have the sampling part here, the decode part, and we get the image here. On the top, you can see that we have the prompt enhancer

09:22 part here as well, which is the the exact prompt we’re enhancing the prompt here, and then we’re generating the text, we’re enhancing the prompt basically, and then we are feeding that prompt back to the actual text encode quant image here, and then the K sampler takes everything here. It takes the model, it takes the positive and negative prompts, it takes the latent image whose noises we’re going to remove. Basically, the latent image is a blank canvas, and the size of the canvas, as we have seen, is directly fed

09:53 from the width and height values on the front page that we’ve seen. So, if I go back, uh you can see that we have the aspect ratio here. 1:1 means the width and height is the same, and it’s fed here. So, once this is ready, uh we can just refresh this and see we still have a few seconds remaining. So, you can see on the top, uh it’s almost 98% here. Just 6 5 to 6 seconds and this will be done. And that’s done. Now, I can do refresh here, and everything seems to be

10:24 fine here. So, let’s go ahead and put in our first prompt. So, let’s go to ChatGPT. Go ahead to the first one, uh exact poster text. So, copy this and just put in the prompt here, and just click on run. So, you can see if you go inside, you can see which step it is in. It is uh inside the conditioning step now, and once it goes to the K sampler, it will start the generation here, and then the VAE decode will translate this back to the image here. So, let’s go back, and I can see that the number of

10:55 jobs which are running here you can just view here, and then you can see this is the job uh 24% total is just 7% is done, and we’ll see how much time does it take for this generation. But, in the meantime, what I can do is I can go to the second prompt here. Go ahead, copy this, and put it here, and just run that as well. So, this will get attached as a second job. I’m going to create some more. I’m going to put in the third job here, as well, and I’m going to say run here, and the fourth job here, as well. Just put it here, and just run this. So,

11:27 the first is done. Now, in order to view this, you can go to assets here, as well, on the left. So, this is midnight signal. The prompt was this one. Uh create a cinematic concert poster midnight signal and life in the city. So, this is really good. So, I just double click here and this will open up here. And that’s really amazing. This is generating on my local system. It’s really amazing. All right, the next one how a bear breeds in winter breeds in winter. This is small organized foot infographics and you can see that slowly heart rate conserve energy weak when conditions

11:59 improve. That’s a really good for quantized model running on our own system. The next was an older SMS woman and if you go here, really great really amazing. Then we have photo studio photograph of a preserve the exact text Northfield. Okay, let’s see. So, we have this water bottle called Northfield. Really great. Let’s go ahead and test the transparency things here. So, let’s copy the prompt and put it here and then run this. Let me take the other prompts as well. Now, if this manual things feel bored to you,

12:30 what you can do is you can always use the APIs of ComfyUI here. So, for using that what you can do is I can just go here. So, right click here and then export as an API here. Now, if you export that as an API, what you will have is this JSON file that you see here. This API and if you open this up, you’ll be able to see that we have this JSON file and and all the things all the selections that we have done is available here and you just need to put the prompts and just use this as

13:00 an API. So, if you’re trying to do batch generation, this is the way out. So, I’m going to put in more prompts here. The next is multiple reference. So, using the supplied portrait reference. So, multiple reference, let me see. Okay, so I’m going to import the other one as well the other JSON file that I downloaded which is the image edit. So, I go to new tab here and then import this image edit here. Okay, so this is a different architecture here, which would need images here, and then you’ll be

13:31 able to edit those images. So, I’m going to see if there is any errors here. So, you can see we have these errors. These are the missing models. So, what I can do is I can download all the missing models here. So, the first thing is that we are missing the model, so it’s downloading the models here, and then we are missing the output as well. All right, let’s go back to our generations that we were having. So, this was the generation. This is the generation of an image. The prompt was something like this, generative full body 3D cartoon

14:01 red panda. This is the one. Amazing. And then we have the create a transparent background asset sheet with exactly four separate objects. The brass compass here, good. Folded paper map, good. Then you have a small red backpack and a silver camping lantern, which is really great. Then we have transparent lettering, exact words make it real, good. So, this is done. Let’s go to the image editing one. Now, in the image editing one, what we need to do is we need to put in reference images. So, we can put in reference images. They use

14:32 as an example of just two images here, but I’m sure you can put in 10 images here that you can see here. So, for the reference image, what I can do is I can put in, let’s see, where does it lead? Okay. Okay, that is for image comparison. So, what you can do is you can put the person here and that image two, we can put in an object here. So, for putting the person, I can just choose and then let’s find person. So, let’s put this person here, and then maybe we can change the clothing here. So, for that, let’s go ahead and get a

15:03 photo of a clothing here. Yep, I just wanted a dress, so I can take this dress. So, so here’s a dress that I’ve selected, so I can go back and just upload that dress. So, choose the file here, go ahead and choose this red dress here, okay? Now, while after the model is completed downloading, then we can just refresh this, and then go ahead and change this and generate a new image here. Okay? So, this is ready. I’ll see in I’ll see you in just a minute. Okay, I guess this is done. Now, I can just do refresh. Cool. Refresh is done.

15:36 So, you can see that I’ve changed this prompt to replace the lady’s white dress with the red dress given, and you can see the utilization on the left. About 95% of the GPU is used, and I’ve changed this to false, the refined prompt. And I can see that K sampler is working here. So, if you double-click here or just single-click, you can see that the K sampler is in process as of now. You can go back and just wait for the output. So, this took about 126 seconds, which is almost double the time and that it takes for a

16:07 text to image or text image generation. And now, we can go ahead to the assets here, and then I can see the output here. So, if you So, you see from the original on the right to the edited on the left, the lady, we see that the dress has been changed. So, in this video, we have seen how we can use Kwench Image 2.1 locally. So, if you’re interested, you can go ahead and check this out and try this out on your own system as well, and let me know how it goes. But, the model that I’ve used is Kwench Image 2.1 uncensored GGGUF from Eben Zerps, and I’ve used the quantized

16:38 4-km version here. So, I’ll see you in another one. Check out the next video.