Qwen-Image 2.1 Hands-On Locally: Prepare a Royal Paan with AI

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: Qwen-Image 2.1 Hands-On Locally: Prepare a Royal Paan with AI · Channel: Fahd Mirza · 15:11 · Watch on YouTube

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

Summary

The video demonstrates Qwen-Image 2.1, a 7-billion-parameter model for text-to-image generation, image editing, and native transparency, using ComfyUI. The presenter shows how to install the model files and explains the roles of the diffusion model, text encoder, and VAE. Tests include transparent images, generating and arranging furniture in a room, assembling paan ingredients into a royal presentation, and producing culturally themed signs and scenes. The video concludes with a prompt-based portrait edit and the presenter’s positive overall assessment.

Key points

Qwen-Image 2.1 is presented as a 7-billion-parameter model with text-to-image, editing, and native transparency capabilities.
The demonstrations use ComfyUI workflows, which the presenter says will be shared on GitHub.
The presenter explains that installation requires diffusion model files, text encoders, and a VAE placed in the appropriate model directories.
A generated transparent phoenix image demonstrates the model’s RGBA output.
The presenter reports VRAM usage of roughly 27–29 GB when the model is fully loaded.
Ten separately generated furniture items are provided as references for a cohesive lounge-room arrangement.
A central test assembles separately generated ingredients into a detailed paan presentation in a Mughal royal-court style.
Rapid-fire tests explore culturally themed imagery from a range of places, with results varying by prompt.
A text prompt edits a portrait by adding a red bridal dress and gold jewelry and removing an object from the subject’s hand.

Chapters

Start	Chapter	Summary
00:01	Introducing Qwen-Image 2.1	The presenter introduces the model and demonstrates text-to-image generation and a transparent image in ComfyUI.
02:09	Model installation and workflow components	The presenter explains where to download and place the model files and describes the diffusion model, text encoder, and VAE.
04:45	Generating and arranging furniture	Ten individual furniture images are generated and then supplied as references for a lounge-room design.
07:22	Assembling a royal paan presentation	The presenter feeds separately generated ingredients into the model and asks it to compose them as a Mughal royal-court paan.
10:29	Rapid-fire cultural image tests	A series of prompts tests signs, streets, and other culturally themed scenes from different places.
14:09	Portrait editing and conclusion	The presenter edits a portrait with a text prompt and gives an overall positive assessment of the model.

Quotes

This is the first and this is I believe one of the salient features of this model. — 01:37

Very quickly it has generated this arrangement and interior design. — 06:19

Now, um I’m not a pan afficionado, but if you are, let me know what do you think about it — 09:59

I think it is pretty good. — 15:11

Concepts and entities

Qwen Image 2.1, ComfyUI, Hugging Face, GitHub, Fahd Mirza, VAE, CFG

Full transcript (original language)

00:01 This image was created with just a text prompt using this new coen image 2.1 model which is a 7 billion parameter model that does text to image editing and native transparency all-in-one model. Let’s run it. I have already got confi installed. As you can see, this is the workflow which I used to create this image from this text prompt which you can see here. A cinematic portrait of an ancient astronaut queen standing on a

00:34 floating obsidian throne. And you can already see the quality is really good. And to be honest after a long time which means couple of months I have seen this good image creation and editing model from any company including flux lab. So let’s see how it goes with further tests. This is Fad Miraa and I welcome you to the channel. I already have Confy UI installed as I mentioned. If you don’t know how to get it installed, just watch this first video which shows you

01:07 to install the Compy UI in as simple way as possible. Let’s do another test in this workflow where I’m going to paste this prompt and run this workflow. While it runs, let me quickly show this prompt which I am using. So this is RGBA image with transparency where we have a majestic phoenix made of fire and molten gold wings fully spread and lot of other detail. Let’s go back and you can see

01:37 that it is already created. Let me open that image. There you go. You see the transparent image. This is the first and this is I believe one of the salient features of this model. This is my VRAM consumption at the moment just under 30 gig which is a standard of these image models lately. 27 to 29 gig of VRAM they consumes when they are fully loaded. The workflows which I’m using I’m using it from comfy

02:09 UI. I have just altered them a bit. I will also upload them in my GitHub repo and we’ll drop the link in video description so that you can also download them. Now, if you are looking to get this thing installed, this is where you need to place your models in this order. And let me quickly show you how you would do that. So, go to this git um hugging face card. From there, just go to files. From files, first go to diffusion models. Download both of

02:41 these. Right click on this downward arrow save link as and then go to models directory wherever you have installed confi UI diffusion models and save there. Then secondly just go back to the files get text encoders and from there all these three files again right click save link as and then go to models and there you will see the text encoder directory.

03:12 Just save all these three files here. Last but not least, you would need to install this VAE or variational autoenccoder. Again, go inside it. This is the smallest file. Save link as. And also you see in the V directory. Just save this file. And that’s about it. After that, you would just go here in the workflows and load your workflow. Give your test prompt here. Hit the run button and it will generate the image.

03:42 That’s about it. Now, diffusion model. Um, very quickly, this is actual image generator. By the way, the brain doing the drawing. Text encoder is what converts your prompt into something the model understands. VAE is what turns the model’s internal representation into the actual pixels. You can see I have done separate videos on all of these. Okay, so that’s out of the way. Let’s do bit more testing on this model. By the way, recently I have

04:13 started this free weekly AI newsletter which you can subscribe at fajhatza.substack.com. Also, if you’re looking to rent a GPU on very good price, you can find the link to mass compute and videos description. And if you want to help out the channel, please become a member. You will find the link in first print comment or at the homepage. Okay, that’s out of the way. Let me do a bit more testing of this model. And now let’s test something more ambitious. We are going to generate 10

04:45 individual furniture pieces from scratch with co image 2.1 model. Then we are going to feed all 10 into this quen as reference images to see if it can arrange them into one cohesive lounge room. Um so let’s test it out. We would need two different workflows. But first I’m going what I’m going to do I am going to generate all these 10 with the workflow I showed you first. So this is the second one which I will use but this is the first

05:16 one. So this is where you can see I’m generating the sofa and you can see how fast this thing is how fast. Anyway so I’m just going to generate all these 10 images and once I’m done I’ll show you then we will arrange them through another workflow. I have generated all these 10 images. As you can see, there is a sofa, armchair, coffee table. There is a rug and then some lamp, bookshelf, a potted plant, wall art, a side table, and then some

05:49 throw rug and pillows. And now I’m going to give it this prompt for the arrangement that arrange these 10 pieces image names whatever the load image node number is. And then I’m asking it to just arrange it cohesively. So let’s see what sort of job it does. So I will just paste this prompt here and run this. Let’s wait for it to finish. And I will just move it away.

06:19 And there you go. Very quickly it has generated this arrangement and interior design. Let me open it in another window. There you go. Look how crisp and clear the whole setup is. Even there are shadows which are correct. Look at the plant. Look at the shadow on the wall. And the bookshelf is there. The sofa is there. Throw rug looks pretty good. Look at the contrast of sunlight and the shadows.

06:50 Even this armchair shadow is quite good. rug looks okay and bit congested but I think it has done really good in a small room. The fiddle leave also look pretty good to me and look at these lines. Pretty nice. Okay, let’s do another test. All right, this new test is something really unique and I can guarantee you no one has ever tested this.

07:22 Nothing you would have seen else where I’m sure about that. Look at this. Pan is a traditional South Asian preparation. A beetle leaf folded around a mix of spices, lime paste, and sweet fillings. Often served in royal or festive settings or even you would find it on the streets of Karachi or Lacnau and any city in Southeast Asia including Dhaka. What we are doing here is generating each individual ingredient

07:54 separately with AI then feeding all 10 into coin image 2.1 asking it to assemble them into one real detailed pan style like it would be presented in a mughal royal court. So you see I have a lot of ingredients here. There is a pawn, there is a soft, there is elite za, there is a ver tutti frutti for garnish. Then we have lot of other things here. And the way I have already generated

08:26 them is with these all prompts right with some soft lighting and all that stuff. Now this is the prompt which I’m going to give to this model and ask it to assemble that for us. So let’s wait for it to see if it is able to do it or not because look interior design all those elements they are quite wellnown must be in the training set of model but this is where the real test of

08:57 model is that because this is not common at all and I’m sure quantium wouldn’t have expected us to generate a pawn out of it. So let’s see how does uh it do with this one. And while it runs, I also want to just touch upon what exactly these settings are in this node. I have done separate videos on them earlier. So look, sampler is the algorithm that actually removes noise step by step to build the image. Scheduler controls the pacing, how much

09:27 noise gets removed at each step. CFG is how strictly the model follows your prompt. And then one here means minimal steering since image 2.1 works best close to that. D type is a precision the model runs at lower precision means faster and lighter on VM at a small quality tradeoff and unit that’s a diffusion model itself the network actually generating the image and meanwhile you can see that it has generated our pawn

09:59 and I have opened it in another window and you know what this is such a royal setting you can see the whole background and offset and then look at all the ingredients which it has placed wonderfully well. Now, um I’m not a pan afficionado, but if you are, let me know what do you think about it because I choose to just eat this sweet stuff from top and that’s it. So, let me know uh in the comments what do you think about this pan.

10:29 Okay, now I am going to do a rapid fire testing. First one, you can see I’m testing out the calligraphy uh from Egypt. And this is a prompt and you can see it has done really good. This miserable jama which I think it has done well but of course I will take your advice if you are from that part of the world. Next up I’m doing a different variety a different style from Bulgaria. So let’s see how this performs. There you go. So this is another cobblestone

11:00 street which is good. But I think this is too way up on here. But again, I wanted it to be a bit more um out of the frame. So I think it has done well. Could be better, but not bad. Let’s do next one from Ulakistan. And again, it’s a ceramic tile mosaic sign at a market entrance. There you go. So this is what it has created. I think it could have done a better job because I don’t see much

11:33 happening here. But anyway, at least it is right according to that culture. Next one. Let’s do from New York. A neon sign above a diner. Let’s see how it goes. It’s quite fast. There you go. This is quite good. What do you think? In a typical New York style. Next up, let’s go to Pinang, Malaysia. Let me run this. Let’s wait. There you go. So, Salamatang

12:05 and all the Southeast Asian plants are there. Beautiful. Okay. Next up, let’s do from Rwanda. And you can read through the prompt. Very nuanced and grounded in the culture. So, there you go. Quite vibrant colors. And the main thing to note down is this border which is right according to their Kali culture. Next up, let’s do one of the most beautiful setting in the world.

12:37 It is the Prague, Czech Republic. Never been there, but have seen that. So, this is what it has created. The engraved reading. I’ll just make it bit bigger. What do you think if you are from Prague? So, this is quite good this uh historic bridge. And I have just zoomed in. Let’s go to Slovakia again. This is a mountain lodge entrance in Tetris.

13:07 There you go. Looks pretty good. Look at the snow. The detail is quite good. Next up, let’s go to Bellarus. It’s a Soviet style metal sign. This facade is in Minsk. There you go. Looks pretty good. Next up, let’s go to Ukraine. And it is in lip. There you go. Beautiful culture. Look at the golden aura in the backdrop. Very

13:37 nice. Next go to Afghanistan. This is a Kapul something in Pashtoi. So let me know what do you think. But the mountain behind this looks quite similar to what we see in Kabul. Let’s go to Belgium from Kabul. And again, another chocolate stuff. There you go. What do you think? Could

14:09 have been better, but I think it looks bit way more natural. And finally, let’s go to Yemen. This is this old city of Sana. A very eye-catching stuff. There you go. So the model is quite good as long as you know the culture is concerned and now I have just generated an enemy and looks pretty good to me. Not bad at all.

14:40 You can also edit the images like just with text prompt. So this is my original image on the left and I gave it this prompt which I’ll quickly show you to put in some red bridal dress elegant gold jewelry and removed the object from her hand. Keep the pose and you can see how good the model is. So it has even enhanced the image a bit. Julie looks good and everything is perfecto. So this is coen image 2.1 for you. Let

15:11 me know what do you think about this model. I think it is pretty good. What do you think? Again, if you want to help out the channel, please become a member. Again, just go to the homepage of channel from there. This is a link. I will also also put it in the first pin comment and subscribe to this newsletter. Thank you for all the support.