Qwen Releases Qwen-Image-2.1-Turbo Open-Weight Model: Test Locally

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 Releases Qwen-Image-2.1-Turbo Open-Weight Model: Test Locally · Channel: Fahd Mirza · 10:30 · Watch on YouTube

Topics: AI content generation, Model inference and optimization, Local and open-source models

Summary

The video tests Qwen-Image-2.1-Turbo, an accelerated version of the image model that generates images in eight denoising steps and can also edit existing images. Across prompts set in several countries, the presenter finds strong results in text rendering, color, atmosphere, and detailed scenes, while noting occasional weaknesses in photorealistic conversion, object shapes, and actions. An editing test preserves the illustration’s outfit and other details while changing hair and shirt colors. The video also explains how prefix KV caching reduces repeated computation to speed up generation.

Key points

Qwen-Image-2.1-Turbo generates images in eight denoising steps and supports both text-to-image generation and image editing.
The presenter tests a low-light Basel Fasnacht scene with masked characters and text on a lantern.
An edit that changes an illustration into a photorealistic portrait in Petra retains some visual identity but remains illustrated.
A follow-up edit changing the shirt and hair colors performs well while preserving the outfit and illustration details.
The Nicaragua workshop image has clear text, visible woven threads, and warm lighting, though the hammock appears as a bundle of threads.
The Chittagong shipyard and Swat Valley scenes are praised for atmosphere, detail, and readable text, with some minor imperfections.
The Ottawa winter scene has convincing lighting and natural faces, but the skaters appear to be walking and the tower does not resemble the cited landmark.
The Ghana scene has vivid colors and readable text, but some objects appear to be jumping at each other.
Prefix KV caching reuses keys and values for unchanged context; only the target image is recomputed at each denoising step.

Chapters

Start	Chapter	Summary
00:01	Introducing Qwen-Image-2.1-Turbo	The presenter introduces the eight-step model and its image-generation and editing capabilities.
01:04	Image editing tests	The presenter tests a portrait edit in Petra, then changes the illustration’s shirt and hair colors.
04:14	Scene generation tests around the world	The video reviews generated scenes from Nicaragua, Bangladesh, Pakistan, and Ottawa, noting strengths and flaws.
08:24	Prefix KV caching and Ghana test	The presenter explains how prefix KV caching speeds generation and reviews a Ghana scene before concluding.

Quotes

What do we get and what do we give up for that speed? — 01:04

So, in terms of editing, you see, it has done wonderfully well. — 03:42

This diagram explains the trick behind the speed called prefix KV caching. — 08:24

I think it’s a strong text detail and color in just eight steps with some trouble on unusual object shapes. — 09:59

Concepts and entities

Qwen-Image-2.1-Turbo, Qwen-Image-2.1, prefix KV caching, Gradio, Fahd Mirza

Full transcript (original language)

00:01 A couple of weeks ago, we looked at Google Image 2.1 on this channel, and today there is a turbo version of it. Which is an accelerated checkpoint of that same model built to generate images in just eight denoising steps. And this is one example where you can see that I have given it a prompt to show me a Basel Fasnacht Morgestraich in Switzerland at 4:00 a.m., a lantern-lit procession of masked carnival cliques on

00:33 wet cobblestone. This prompt is testing readable text on a lantern, which is quite hard to render, plus difficult low light with glowing sources and many detailed masked characters in one scene, and the model has thrown out of the park. This is a VRAM consumption of this model at the moment. Now, another cool thing about this model is that not only it can generate images, it can also edit them. This one keeps the same seven seven

01:04 billion architecture and handles both text to image and image editing in place. So, the question for today is simple. What do we get and what do we give up for that speed? Let’s generate our editing test now. For this one, you can see that the quality is quite good. Let me do the editing one. Let me upload an image from my local system. So, I have uploaded this image from my local system, which is an illustrated rendering of my own image, and you can

01:35 see the prompt. So, this one is testing whether the edit keeps identity, outfit, and colors while changing the style. And you can see that I have also asked it to just take me to Jordan, just inside some canyon at Petra, and also just convert this illustration into a full-blown photorealistic portrait. By the way, this is Fahad Mirza and I welcome you to the channel. If you want to help out the channel, please become a member. The link is in first pinned comment and

02:06 at the home page. Also, please subscribe to this newsletter, which is free at fahadmirza.substack.com. And you can see that this is my GPU, which I’m using. If you’re looking to rent a GPU on very good price, you can find the link to my computing videos description. Let’s go back and check this one out. Now, you can see that it has put the canyon right. It has also given some of the Jordanian architecture, but the right-hand side

02:38 image is not that photorealistic. If you compare it to the previous or the actual coin image 2.1, over there, it was quite good in terms of photorealism. But, this is where it failed. It has tried, especially if you look at the hair and eyes, but other than that, you know, it is still an illustration without any doubt. Let’s do another change. I’ll just ask it to change the shirt color to red and hair color to maybe blonde.

03:11 Let’s see what sort of uh image it produces. And this is just a simple Gradio stuff, which I’m running. I will drop the link to their model card in video’s description, which has all the installation instruction. And I have just put in a Gradio demo on top of it. If you really want to see how to get it installed, just go to my channel, search with this coin image 2.1. It’s the same instruction, just replace the model ID, this one, with that, without turbo, and you should be good to go. Okay, let’s

03:42 wait for it to get generated. There you go. So, in terms of editing, you see, it has done wonderfully well. Really, really good. It has maintained the outfit and all the illustration. The hair looks great. Plus the shirt color also looks great. And the bowl and the uh you know, all the creases that you know, a hint of jowl and everything looks really cool. Even if you look at the curves of the hair, they are pretty good. Pretty good.

04:14 The eyebrows are okay. Yeah, I’m happy with this result. What do you think? And now let’s go to Nicaragua. Look at this prompt. It is showing a Nicaraguan artisan workshop in Masaya with a hammock being woven. And it is testing close-up texture on thread and hands text on a sign and warm indoor light through the windows. So, it is very clean. The text rendering is great. The weaver’s lined hands and

04:44 braids are sharp. And the woven threads on the loom are clearly visible. The cat, the coffee cup with steam, and the window light stripes all made it in. Um but you know, the only thing is that the hammock on the wall came out as a hanging thread bundle rather than a finished hammock, but still I think not bad at all. What do you think? Next we are going to Chittagong, Bangladesh and I think this is one of the best image generated I have seen.

05:16 Um the prompt is showing a Sitakunda ship breaking yard near Chittagong at sunset. It tests massive scale, rusted metal texture, and readable text on a ship hull. If you look at the image, the scale and mood landed well. The ship towers over the workers. The rust gradient and the sparks at the base look quite convincing and the sunset sky, smoke, red sail sailboat reflections in the tidal pool, and the dog in the corner all made it

05:48 in. And the text is readable on the hull, but it appears twice as you can see, which is fine. Not I think on either side. Um it is fine. That is fine. Um the workers are also okay. Um I think it is the whole atmosphere, the synergy is landed quite well. So, I think I will take it as is. Pretty cool. And look at this one. This is another masterpiece, I would say. The prompt on

06:18 the left shows Swat Valley apple orchard village in Pakistan in autumn. It tests faces and hands in close detail, layered mountain depth and text on a hand-painted truck. The truck text Swat ka Shahzada is spelled correctly and the ornate floral border looks quite authentic. The farmer’s face, beard, and Pakol-style cap are all detailed and the hand passing the cup is quite clean. The mountain layers, pink peaks, stream

06:49 autumn leaves, and wooden balcony house um they all look quite good to me. But there are a few weak spots as well. Like the apple branches are mostly foliage with only a few visible, but still um I think because it’s a side one, so we will take it as is. But other than that, I think it is pretty cool. Uh it has you know, depicted the culture very well. And here is another good one. We are in Ottawa, Canada. It’s a canal skateway on

07:20 a winter night. It is testing snow and ice texture, glowing light sources, and text on the signboard again. And again, the text is really good. Sits nicely in the lamplight. Then there is a Parliament Hill clock tower, the glowing turquoise ice beer, the steam from the vendor’s counter, and the cold blue versus warm amber lighting all is looking really cool. The family faces are also natural. You don’t have to look closely and the cup of hot chocolate is clear, too.

07:52 Um there are few, you know, spots you know, weak spots I would say. Like the skaters are mostly walking in the crowd instead of actually skating or gliding, I would say. And I think the tower also doesn’t look like the Chateau Laurier which I saw when I was in Ottawa about 10 years ago. It is more like a Gothic tower. Looks cool, no doubt about that, but I think doesn’t look like anything like the actual one. Let’s do one final test. I’m just going

08:24 to take you to Ghana, Africa. I’ll just run it and I wanted to show you something while it works through it. Have a look at this one. So, this diagram explains the trick behind the speed called prefix KV caching. The model splits its work into two parts. The system prefix, the optional input image and the edit instructions never change while the image is being generated. So, their keys and values are computed once and reused.

08:54 The blue blocks labeled compute once reused KV only the target image, the green block, is recomputed at each denoising step. So, with just eight steps the model skips a lot of repeated work and that is why generation is fast. Especially for editing where the input image is the heaviest context. Okay, so there you go and you can see that this looks pretty cool and so much happening in this small

09:27 sort of frame. Text is readable. Rooster’s glossy feathers are there. Apron is okay. The Kente cloth from Ghana and the paint cans are all sharp. Colors are vivid, too. But again, um you can see that you know, it could be better. Like, I mean, they’re all sort of jumping at each other. Um the detail and colors are okay, but I think it could have been better, I would say. And this is, you know, fantasy coffins, by the way. It’s a very

09:59 interesting concept. If you want to read more about it, uh maybe just go and have a search at it. It’s very, very interesting topic out there. Anyway. So, I think, uh Ganime 2.1 Turbo, I think it’s a strong text detail and color in just eight steps with some trouble on unusual object shapes. So, thanks for watching. Let me know what do you think. And again, if you want to help out, please become a member. Just go to the home page and then from there, just click on this link. Thank you for all the

10:30 support.