00:00So, Kwen just released Kwen Image 2.1, and the number that stood out to me is seven

00:057 billion parameters

00:08That is small

00:08The visual generation part is 32 single stream DIT layers, 7 billion parameters, and it is doing generation, editing, transparency, and multi-reference composition all in one model

00:21One text encoder, native 2K output

00:25Here is the part I did not expect

00:25Back in December, Kwen shipped a full separate model just to handle transparent images, Kwen Image Layered

00:35That capability is now built straight into this one

00:36You write your prompt, and it hands you an RGBA image with a real alpha channel

00:41No background removal node, no rembg, no masking

00:44And it works backwards, too

00:47You give it an ordinary photograph, and it pulls the subject out as a transparent layer you can drop into anything

00:55Then the editing side

00:55It takes up to 10 reference images

00:5810

00:58You can hand it a model, an outfit, shoes, a bag, a hat, and it builds the full look in one photograph

01:06You can circle a region, or paint on the image, or feed it a separate mask, and it only touches that area

01:13Faces and product details hold up much better than before, and all of this runs natively in ComfyUI on day zero

01:18No custom nodes required

01:21The int8 build is about 16 GB for the diffusion model and the text encoder together

01:26A 16 GB card handles that fine, and it will still run below that with offloading

01:34One more thing

01:34I built a prompt enhancer node for ComfyUI that wires up the official prompt rewriting models

01:42So, we are going to set it up and see what breaks

01:46Okay, so in ComfyUI, this is something I am currently testing, and it is practical for e-commerce

01:50Product demos, a showcase of whatever clothing, accessories, or physical product you are selling online, Shopify, whatever it is

02:00You can have your model, and now it comes out with a transparent background right out of the box

02:04The generated image kept the pose of the character

02:08Then I can use the image edit functions to add those accessories, colors, clothing, all kinds of stuff onto the character

02:13That is a game-changer inside a 7 billion parameter model

02:18You could get something like this from Nano Banana a few months ago, and a lot of advertising agencies were told to use that

02:25So, first of all, let me take a look at what we have in the Hugging Face repo

02:29In here, we have the diffusion model, the text encoder, and the VAE

02:32You download each of those

02:36The model weights come in BF16, and there is int8 conv rot plus w4a8 for the Qwen-3-VL text encoder for GPU with even less memory

02:47Image 2.1 is easy to run without any LoRA plugin or additional things, but the prompting is tricky sometimes

02:54Before I downloaded this model, I downloaded the full model weights and prepare for LoRA training

03:00But without LoRA training, go first to the GitHub repo of Qwen Image 2.1

03:03You will see hints called the prompt rewrite

03:07That helps you use something like a prompt enhancer to rewrite the prompt

03:11So, based on the documentation in there and the examples in the source code, I created the prompt enhancer ComfyUI custom node using the same mechanism and logic

03:20It takes what is on the official Qwen Image GitHub and turns it into a ComfyUI custom node

03:27Now, back to the graph

03:27Before we test this graph, let me show the official ComfyUI workflow

03:31This is the official ComfyUI example workflow that they were showing

03:36I tested this one first, and then I optimized what I wanted and customized the data flow I needed into another ComfyUI workflow

03:43This one is beginner-friendly

03:46It is all compacted into one single subgraph where you choose the model For example, we have Qwen image 2.1 for the diffusion model and Qwen 3 VL

03:55I have FP8 and BF16

04:02For this purpose, I will most likely use BF16 because it is a small model, so it does not matter to me

04:10Go to the VAE

04:10There is the Qwen image 2.1 VAE in this loader

04:12Follow all three download steps in the ComfyUI official repositories in this Hugging Face repo

04:20Going back to the custom nodes, the text prompts are pre-filled in this example workflow

04:26What you do is generate once you download all three files, select them accordingly, select your aspect ratios, and click and run

04:31I am not going to save this image

04:33In this demo, I will use a preview

04:35Let's run it

04:45This last generated image, 5 seconds, 25 steps

04:48That is all it took

04:48I do not think this model needs a turbo LoRA or a low sampling steps LoRA because even if I turn it up to 50 steps, it is still fast

04:57But you do need enough memory on your GPU to get this speed

05:00Even at 50 steps, 10 seconds, no problem

05:06A lower memory GPU might take a little longer, but not obviously longer

05:08I believe within 1 minute you can get an image

05:11Generation speed is not a problem with this model

05:14This is what you get by default from the ComfyUI example workflow when you download and run text to image

05:21You will see the texture

05:23Although the texture is not very sharp because of the graphic style, a lot of the details are clear

05:27For a 7 billion parameter model, this does pretty well for its size

05:33The second example is image edit

05:33This model does text to image and image edit

05:38In this example, I have a character with some outfit layered in the image and I changed a little of the default text prompt to put the jacket and the dress from image two onto the character, preserving all the original facial features, hair, body shape, pose, and the background from image one

05:55There, we get the after-generated result of this image edit

05:57The pattern of the clothing is something I always look at when we talk about Tryon

06:01Zoom in here

06:04Those little flowers are replicated in the output image

06:06Those colorations from the dress to the output and the jacket texture pattern

06:10Everything looks very similar to what is there already

06:12So, it is kind of ready for e-commerce owners, Shopify stores, to run product image generation locally, customized, or even to remove the background, which is something important that I have done a lot before

06:27Now, my own customized workflow for this model

06:29Not fancy, it is one-pass sampling for both types of image generation

06:33On the top, here is text-to-image

06:35I added my own custom nodes for prompt enhance or prompt rewrite

06:42Sometimes I do not need a long-form prompt

06:43What it should be called is the formatted text prompt that the Kwen image models understand

06:50It is not only about generating an essay-long prompt

06:52Sometimes you need a formatted prompt so the model understands what is going to be generated

07:00This is why the system prompt from the Kwen team matters and why they put one in for both text-to-image and image edit

07:06There is a structure to it

07:08Quality image output does not equal a long-form essay of text prompts

07:13Make things efficient, not bulky, like a large chunk of essay text

07:15We have the text-to-image module

07:17First of all, let's run this and make sure it works

07:22Nowadays, I do not think I need to demonstrate realism-style images or characters standing right in front of you

07:29I want to go for harder, more logical things

07:30For example, something like this

07:33An automatic watch

07:33What is inside the watch shell? It has some parts, the wheels, gears, whatever is spinning, showing the assembly steps of the automatic watch, what goes in there as a component

07:44This shows the model can do something complex

07:49Another example, classical and quantum computing, an easy explainer of the concept, something clear with text able to render in the image

07:57This one looks like a whiteboard with words and text written in different colors

08:03There is something wrong in here, in this sign, but the spelling of the text is kind of okay

08:09This is one example I want to show that within the text prompt you can create something like this

08:16Another one is a screenshot of a trading terminal

08:18Notice I am using 50 steps here

08:20I push higher steps because this model is fast, so who cares? I just use 50 steps

08:27This is the trading stock market thing

08:29As you can see, some marks in here are not accurate in terms of trading using the candle graph

08:35The candle graph for uptrend and downtrend has the right color, but the one at the bottom, the RSI here, is not the accurate pinpoint of this arrow

08:42But looking at the whole image, it does catch the general look of a trading stock market graph

08:50Another example, a chalkboard with a syntax tree writing on the board

08:52The color looks like a chalkboard color, but the text font is a little bit of a computer font, not kind of a handwriting font

09:02It tried to get the idea across

09:05Another image I tried is a zoning map, a blueprint of construction buildings

09:07We have residential area, commercial, industrial area, and the train underground with some numbers, but sometimes they do not make sense, especially on the industrial buildings

09:20For industrial, we have 3 m within this width, and above that is 12 m

09:23That does not make sense for these numbers, and these numbers do not make sense either

09:30It is trying to present it so it looks like it makes sense

09:33But if you go deep and look at those numbers, especially those of us with mathematics and computer backgrounds, we always look at numbers

09:39And when we see this kind of thing, we can prove it is wrong

09:43It is trying to give you the concept, the layout, and the idea of it

09:48But a lot of numbers and text, you have to adjust by yourself

09:50You cannot just one prompt generate everything with this image model

09:56As for realism styles, which a lot of people like, yes, you can do that with this model

10:00And I assume it is kind of okay with one sampling pass through all these evaluations of how it runs with one sampling

10:07I also want to try Asian characters because Kwai Image previously had heavy plastic skin for Asian characters

10:14In this text prompt here, I am trying to see if it still happens with this 2.1 version

10:20There we get the generated output

10:22The face skin is kind of okay

10:22Although, you know, it is a one-pass image

10:27We are only looking at whether there is any plastic skin

10:29And it is not too obvious in some ways

10:31The fingers, everything, five fingers here

10:36Let's take a look at the image edit

10:39This is more exciting, more fun to play with than the text to image

10:41Here I have a character in some outfits

10:43I try using a tag like image one, image two in this style

10:49This is what the Kwai documentation tells you, and using that to swap the character's outfit, especially the handbag and the high-heel shoes, I want them held in the hand

10:57Let's see what we have in the output

11:02It retains, preserves the positions of the character pose

11:04The good thing is that this model can generate transparent backgrounds out of the box

11:10This is essential

11:10We can play around with multiple image references

11:12Like this, where I resize all the input images here, four of them

11:16Notice Kwai Image 2.1 can take up to 10 image references for image edit, and I just take four right now for a general try-on demo

11:26We got these pants, a shirt, and sunglasses to put onto this character

11:33Here, we will be swapping out all those and putting them in

11:35Let's bring this output here together with all of that and see how that looks

11:40The sunglasses shape is almost the same as the shirt pattern and the button on the chest

11:44Yep, pretty much the same one, especially the inner layer of the neck here, the same color

11:52The pants here are a little bit awkward

11:54There are two straps hanging out untucked, and the AI still replicated those two strings untucked

12:01But I randomly picked it from Pinterest and it just works at replicating this style

12:06So, it can make multiple images pretty accurately, especially Kwen from Alibaba

12:09Of course, they will train their image model pretty well for clothing outfit try-on for e-commerce business

12:16The next example is interior design

12:19Every family might move, and you want to decorate your new house, an empty living room like this

12:24You want to search online, grab those images from Pinterest, whatever it is, mix and match the ideas, and fill in this empty room

12:32This idea was shown in the Kwen image blog post as well

12:36It can take up to 10 images as reference

12:39I am trying five since this is a minimalist style room that I want to build

12:43Let's run this and see how that goes

12:46In the text prompt here, all I did is say that image one is the empty living room, and then all the way down here, all the stuff is the layout and placement of whichever objects and put that into the image

12:56Let's see what we have in the output

13:00The result is something like this

13:03Most of the things coming from the reference are kind of right for the objects, but sometimes the size is not exact because there is no spatial measurement in an image model like Kwen image

13:12So, the empty room, the coffee table here, I picked this coffee table on purpose because it is not a normal coffee table shape, and I want to see if the AI model can generate this

13:22It works

13:25The colorations match the environment's lighting

13:27The sofa and the love seat, two of those, one came out a bit off-color, kind of green in the generated image

13:35Moving on to the TV wall here

13:37Yeah, I try to put this TV wall style and all the decorations into this living room, and it works here

13:41Although the sizing is not going to be exactly one-to-one, it is trying to use it for interior design, the concept, how to mock up the layout

13:49It helps regular people who are moving house

13:52If you're talking about a professional designer, this is not the tool you want for chasing perfection, especially when you need exact measurements

14:01So far, this model did pretty good for overall general-purpose image generation

14:05Although they already released Kwench 3.0 on the Kwench image or the Kwench Studio website, this one still ranks pretty high on the benchmark against the current market image generation models

14:17Something worth mentioning, this 7 billion parameter model just crossed Nano Banana 2.0 in the benchmarking, just one point here

14:26Kind of cool

14:26Nice to reference for the benchmarking, but the numbers are not always exactly what we follow when we play with it

14:33Sometimes we have different requirements, different expectations for the specifications you want to achieve for your generated output

14:38There are two groups here, two nodes, that cover the things we want as one image generation model combined with image edit, and it is very lightweight at 7 billion parameters

14:50The next test is different camera angles

14:51We have an image like this

14:54Let's remove the other things we do not need, so they do not block our view here

14:58We are going to change it into different camera angles

15:02Here I am trying different camera angles with this image and seeing how that looks

15:05But a camera looking from outside the window, coming back in, does not work because there is no information about the character's face in the reference image

15:13Other than that, the other text prompts for different camera angles are working

15:19This one is a top-down view of the room, able to switch the camera to other angles like this, and it works

15:25Let's see if I provide the character's face, whether it can do those camera view rotations

15:29If I provide a character sheet like this, I am actually able to use that angle of the opposite side of the image, because now we have the character information, what the front view of this character is, and the AI is able to predict what is in the room looking from the opposite side

15:43So, sometimes we need to test with different input information

15:47This is not about being artistic

15:49In general, this is just providing enough data for the AI model to understand and recognize it

15:56So far, this text prompt for the opposite side works, even with different camera angles

16:00But sometimes the angles are still wrong in some generations, like the bed and the pillow here, the windows should not be at this angle

16:08There is still something it cannot predict from the room I gave it, but it is good that it proves the concept

16:12Once I provide the character front view with the character sheet, it is able to get the character at least right here

16:18The next step would be to have another angle of this room, then you get the right angle that makes sense

16:23Right now, it still does not make sense, but as proof of concept for half the character face, it works

16:30Another task

16:30I am going to add this new outfit for this character and put it into this room, making another angle for this character

16:39This was kind of difficult back in the day for older image generation models, having multiple tasks like this in one go

16:46Right now, we can do that with this character, with this clothing in the right outfit

16:50The positions here, those decorations in the room, remain

16:53Pretty cool

16:57That is something we can do with this image generation model

16:58We'll test more from it and see you.