Qwen 3.5 in ComfyUI + Align Tool & Pixaroma Nodes Updates (Ep16)

Añadido: 2026-05-12

[00:00:00]In today's episode, I show you how to get prompts from images or text using the Quen 3.5 model, plus a lot of updates for Pixarroma nodes and a new align tool. So, make sure you watch until the end. If you use easy installer, make sure you update it using this BAT file. That way, you will get the latest Pixaroma nodes. If you are using a different Comfy UI version, make sure you update the nodes from manager to get the updates. Speaking of updates, IVO added a few more options to the easy installer. From this update forward, you can update everything from this menu, both easy installer and also Comfy UI or Comfy UI plus nodes. Then you can start Comfy UI either in the browser or the desktop version. Thanks to Ivo for the updates. Now in Comfy UI, if you don't have easy installer, you can go to manager, then custom nodes manager, and search for Pixarroma. You can update it from here or switch to older versions if needed. You will also see new versions as I continue to add more nodes. If installing from manager does not work because the node is too new, you can always use the git command to clone it manually from GitHub. Also on GitHub and Discord, you have the workflows I use today. So let's start with this Flux Klein workflow. The note contains everything you need to run the workflows and I will show you more updates I did for this note node later. So make sure you download the models using the buttons and each model has the folder location where you need to place it.

[00:01:36]Then you need to select the models in the nodes. It is easy to see which model goes to which node if you check the title. After downloading, you need to press R to refresh the node definitions so the nodes can see them. If we run this workflow, it is a simple text to image workflow like in episode 4, just with some new nodes added to make it easier to use. I did some updates to the preview node. So now you can choose how the node acts, either like a preview image node or like a save image node. So if you want, you can leave it on preview and only save the images you want. If I click on this button, it will save it in the output folder. You also have the option to save it to any other folder like you usually save images from the web. So that is the update for the preview node. Moving to the resolution pixoma node, I added a few more resolutions, so it should be enough to cover most use cases like Instagram ratios and so on while still using multiples of eight to stay compatible with AI models. For those who want a ratio that is not in the list, now you can add it yourself using custom ratio.

[00:02:44]So I can add maybe a 1:2 ratio and run the workflow to get that ratio or I can use the switch button to swap the sizes.

[00:02:52]And now I get a 2:1 ratio. Of course, you also have custom resolution, so you can enter any pixel size there, but I added some limits, no bigger than 4K. So I hope this node helps. Let me make some room here to show you an update I did to the image crop node. Now we can input an image and crop it. Or if nothing is connected, you can select the node and use Ctrl +V to paste an image there, then crop it after. You can also have two preview image nodes, one for the normal image and one for a cropped version of the image if you want. For example, since the generated image has eight extra pixels, I can use a crop to get the exact size after it is generated. While recording, I noticed that my preview image node did not show the new size there. So, I fixed it. And by the time you watch this video, it should show the new size and the image correctly. The cool part of the crop node is that I can open the node and crop visually, like dragging the corners and moving the crop where I want. I can even align that crop. So, it gives you much more control. No more guessing what it will crop. When you save it, those settings are updated in the node. I can also adjust it directly from the node if I do not want to open the crop window, but it has fewer options and is less intuitive. Now, when I run the workflow, I will also get a cropped version like you see here. It also depends on the canvas size. I usually use the settings I put as default with center crop. So when I run it, I can get an exact image size cropped from the center. That way I can adjust the ratio and get another center crop automatically. So it can be useful in many cases when you need an exact size and the AI generates slightly different sizes, usually with a few pixels difference. Let me show you how you can use Quen 3.5 to generate prompts. If you have enough VRAMm, you can later connect it to a different image model workflow and generate both the prompt and the image with it. Even though Comfy UI loads it through a clip style node, Quen 3.5 is actually a vision language LLM capable of generating full text responses from images or prompts. So you download this model and place it in the text encoders folder. After downloading, press R, then select it in the node. To make it easier, I used a concatenate node that has two fields. So you can keep the formula in one field and put extra text as needed in the second field. Then this node generates the text. The show text preview node was also updated to allow copying the content, to have text output support, and to fix some resize problems. Basically, this workflow uses text to generate another text or prompt, like a more detailed version, as you can see here. Now, if you try to run it a second time, nothing happens because you did not change the formula or text. So, if you want to keep everything the same and generate a new prompt, you need to change the seed manually from here. I assume it works like this so that when it is inside a workflow, it generates the prompt once and the next time you run the workflow, it keeps using the same prompt again and again to generate different images until you decide you want a different prompt. So if I change the text here and run it again, you will see that it works again because it detected a change even if the seed stayed the same. Now let's see how we can use this in any workflow assuming you have enough VRAMm to run two models in the same workflow. So I open this Flux Klein workflow, the one we used earlier. Then I go to the Quen workflow, select the nodes and copy them. After that I paste them here. I already prepared a workflow for you, but I just wanted to show you how you can connect different workflows together. So, we need to get this prompt into the text encoder of the Flux Client workflow. So, we can continue from here using the generated text. When you move near the top left corner of the text field, you get a dot where you can connect it. If you do not want the preview, you can also get the text directly from this text generate node. I personally like connecting it through the preview because it is easier to follow the workflow flow. So it starts from the Quen model. It generates the prompt and that prompt is now used inside the flux workflow. Our workflow is now bigger and takes more time. But it is also smarter.

[00:07:23]Let me test it with a cartoon bunny. As you can see, we start with a simple prompt here. Then it generates a long detailed prompt here. And after that the prompt goes into the text encoder. Then the AI does its magic in K sampler and we got our image generated. Pretty cool, right? If you run the workflow again, you can see that it does not use the Quen model again. So now it is much faster the second time. For example, I can change ratios and in a few seconds I get a new image. I have 24 GB of VRAM and it is quite fast for me. The quality is also pretty good when it does not mess up hands and other details. You can always generate multiple images until you get one that looks right. But wait, there is more. We can use the Quen model to generate a prompt from an image. I know it can be confusing for many beginners because there are so many Quen models. Some are Quen image, some are Quenedit, some Quen models only generate text and read images and so on. If we remove this load image node, it would become the previous texttop prompt workflow we had before. So the difference is this load image node connected to the text generate node plus the formula is different because now it describes the image instead of generating a prompt from text. So it looks at the image and generates a prompt from that image. But it is better to see it with an actual example. If I open this workflow, I already attached the Quinn workflow to the Flux Klein workflow. And you can attach it the same way to any other workflow that has a text encoder. This one is connected here like before. You can connect it however you want as long as it outputs text.

[00:09:07]Then you can decide what resolution you want and hit generate. Now the Quen model looked at the image and generated a prompt description of what it saw. And now we got this bunny image. It also depends on the model and how it was trained. For example, the bunny image was originally generated using a Quen image model a few months back. So, if I attached it to the same model that generated the original image, it would probably look even more similar. Here we have the formula, but the cool part is that you can add more instructions. For example, I want to adjust the prompt so when it describes the image, it uses a panda instead of a bunny. Now, the prompt describes a panda, and the final generated image is a panda character dressed in similar clothes. I think it came out quite well considering I did not use image to image or image reference. I only used a prompt generated by Quen. So, I can go and change the image. Now, let me maybe put a similar ratio first and then generate.

[00:10:10]Now, I got a panda in a similar color and pose to the original bunny. So adjust the formula to get better results and try a few seeds. Let me try with this vase for example. I will remove the extra prompt and now I got a vase of flowers. But maybe I get a more similar result if I use the same ratio. The purpose is not to get the exact image but to get better prompts that we can use and adjust. So let me adjust this one. Maybe to get a watercolor painting instead. To get better results, you probably need to change the original formula, too. But many times it works okay as it is.

[00:10:51]Here are a few more examples using images from Google and the prompts Quen generated for me. And just to show how important the image model is, here is the same prompt used with Nano Banana 2 since that one can create more realistic results than Flux Klein.

[00:11:11]Let's move to the next node update and I will show you in a new workflow. This one took a few days to fix bugs and add more things to it. Even though it looks simple, every new feature I added introduced more bugs. Then I got stuck in a loop with bug after bug until I got it to work or changed the approach and asked it in a different way. First of all, the icons had a lot of bugs because they were sharing color with other elements. Now we can have icons change their color and size and insert those icons into the note. You can delete one if you do not like it and add another one with a different size or color. So it should be quite easy to use now and useful for making the notes more interesting. I used Nano Banana 2 and Adobe Illustrator to create those SVG vector icons. So I hope they are useful.

[00:12:02]For the note background, it is easy to change it to any swatch color. To avoid more bugs, I made it apply when selected. If you need more colors, you can get them from here. It works kind of like Photoshop where you select the color you want or add the color code so you can make your notes look how you like. I also changed how buttons work.

[00:12:22]Now you can select the button type from here, change the text on the button, add a link, and enable or disable the size hint for when you have models. Then you can insert nice buttons that people can click to download the models. Another useful thing is that you can edit the button after there is a little pencil icon and then you can change it to a different color, change the icon and update the button. So all my workflows will now have that note to make it easier for you to get everything you need for a workflow. Then you can add the folder location suggestion for the models. If I use this button, you can see it will insert a folder followed by the comfy UI model location. You can add it in any color you want. Then change the path that suggests the folder location for the models. I found it quite useful and fast to add inside the note. Then I did the same for the grid.

[00:13:16]Now it has its own colors and I can change the color of the header too if it is enabled. Then you can type inside that grid easily, similar to how you type in Excel. After you save, you need to adjust the note so it fits whatever you are trying to place inside the note.

[00:13:32]I like to resize it so it does not show a scroll bar on the side. You can change the title so it really becomes a personalized note. You can also change the color from here, but I recommend using the settings for that to get a custom color that looks nice and so you can see what the background looks like when you add your text. For more help with the note node, you can check the help button that explains what each icon does. So, make sure you check that. You can also see how the code is used if you prefer adding it that way, but it should be much easier visually. If you decide to use code, you can see how it looks here. And you can even create a custom GPT that generates that format for you.

[00:14:13]I added a text formula inside the custom nodes folder in the assets folder that you can use with chat GPT or Gemini to generate that format. Then you can preview it and save the note. And that is all for the note node. Now let me show you a new node that is more front-end based. It is not something you add to the workflow. It is something you activate from the interface. You should see this icon on the top bar near the manager. That is the align pixaroma tool. If you click on it, it becomes activated and turns orange. Now when I move a node, I get little guidelines similar to Photoshop and other software.

[00:14:53]So I can see when the edge of a node is aligned with another node. You can also zoom in for more precise control. And you will get a little snap and that orange line when it is aligned. The same thing works for the center of the node and also when resizing a node. For example, if I select the corner of this K sampler node and drag it, you can see the guides when it aligns with other nodes. Again, if there are too many nodes, it might be harder to align, so zooming in helps a lot, but I think it is quite a useful tool. I usually keep it active only when I want to align nodes. The rest of the time, I keep it disabled so it does not keep flashing non-stop. I also added some more settings in the settings menu. Look for Pixaroma there and you will see that you can turn the align tool on and off from the settings as well. You also have some settings for snap distance which depends on the screen, but usually it works fine with the default settings and I just zoom in for more control. If the align tool is active and you want to bypass it temporarily, you can hold shift and it will ignore the align tool until you release the shift key. So play around with it and see if it helps and turn it off when you are not using it. I also added a small update to the image composer. Let me add a node really quick. Then upload an image.

[00:16:16]In the transform tool, you now also have a blur slider. Some people asked for this blur option, so I hope it is helpful. Make sure you check the GitHub page for Pixarroma nodes. Maybe give it a star if you like the project. Scroll down until you find the change log. And usually there you can see what bugs I fixed and what new nodes I added. It is really a lot of work. I do all of this with claw desktop and sometimes it can be difficult if you do not know coding because you need to explain to the AI exactly what you want. It can also use different terms for some things and misunderstand what you mean. Sometimes I spend half a day trying to fix a bug that is so deep that it is hard even for the AI to find it. So after every new thing I add, I have to test it, then explain what needs to change and test it again. Sometimes it even thinks for 10 minutes trying to figure out how to fix a bug. So it is quite timeconuming. But I hope these nodes help you. I forgot to show you something in the preview image node. If you generate multiple images at once in batches, keep in mind that it will need more VRAMm for that. So if I want to generate three images and let me run the workflow, I can see all three images inside the node. You can set it before you run to preview from here.

[00:17:36]Then click on an image to see it larger.

[00:17:38]And you can close it from here as well.

[00:17:41]So if you like image three, you can save only that image using the buttons if you want. Also, while you have an image selected, if you click on it again or use the left and right arrow keys, you can navigate to the next image. That is all. Now, I need some sleep.

[00:18:00]That is all for today. Thank you legends and everyone who subscribed to the membership.

[00:18:08]I hope you enjoyed today's video. Leave a like and a comment to help with the YouTube algorithm.

[00:18:17]Have a great day.