The training dataset consisted of artwork where each pixel was equal to 8x8. This is important because if you need "perfect" pixels, I recommend RESIZE at 0.125 (If you are repelled by the factor, or by 12.5% if you have percentage manipulations), and then RESIZE at 8.000, so that the pixels acquire their true nature. Use the Nearest Neighbor method when resizing.

Changelog

Version 8.0

Trigger words (place at prompt start):
masterpiece, pixpix, 8-bit, pixel_art

Model: waiillustriousSDXL_v160
CFG: 3.5 | Steps: 30 | Sampler: dpmpp_2m | Scheduler: Karras

Recommended: using custom node in ComfyUI—I believe this is the right approach. How to use my custom node

VAE note

The Pixelate-x8 VAE (https://civarchive.com/models/1492329) may conflict with this LoRA: backgrounds occasionally fragment into gray noise, saturation drops.

All 92 training images were upscaled to 1984×1984 by extending canvas with nearest-background colour (no content distortion). Without this uniformity and proper tagging, the LoRA absorbed ≈20% of the intended style. It works - but it's a workaround, not a solution.

Version 8.0 - pixel-art style with tag-triggered sprite composition.
Use training tags to force sprite framing. It won't always be easy, I'm working on it.
Without them - clean pixel-art in the same style, but free composition.
Not a sprite generator. A style LoRA that can produce sprites when prompted correctly.

Version 7.05 🔨

Usage Recommendations:

I still recommend using custom node in ComfyUI—I believe this is the right approach. How to use my custom node

Model: illustrij_v18 | minij_e4 | Plant Milk Model Suite Walnut | indexed v1 prototypeIllustrious
CFG: 3–4
Steps: 28+
Sampler: Euler | EulerA
Scheduler: Simple | sgm_uniform | Karras

Write your prompt after my trigger words: masterpiece, pixpix, 8-bit, pixel_art

VAE:

I usually used baked VAE, but after updating Comfy, I don’t have a choice but to use the first one available, which happened to be lunaXLILNAIVAE_luna. Even if I wanted to, I can’t check if it makes any difference or not.

Speaking of VAE: I recently came across an SDXL-based VAE, and it’s pixelated. The results look amazing, but SDXL itself is outdated, and that VAE conflicts with my LoRAs. I’d love to train my own VAE—imagine, about six people offered to help with their hardware, but something always went wrong, and I still haven’t been able to do anything on external equipment.

I’m certain that a VAE trained on my data could produce better results than a LoRA.

I haven’t been able to get back to this for a while. Laziness has been choking me.

Test Model:

I’m using 72 images (none of which were in previous training, so there should definitely be a difference compared to other versions).

I like that this version produces interesting horizontal landscapes and unusual characters. The girls sometimes have issues with their eyes, but I know why—it’s clearly a dataset issue. Fixing it would require creating a whole new model.

Over the past three months, I’ve been hit with some serious apathy, but I’ve come to understand some technical truths. For example: did you know that if you allow even one questionable-quality image in your dataset, all the other flawless images will have almost no positive effect? The bad drags down the good. A neural network never forgets what it’s seen, and if there’s something questionable, it will appear in every generation. So even if we have what seem like good images, we’ll still end up with mediocre results because some things—intentionally drawn by artists—might turn out to be artifacts that appear in every generation.

The larger the dataset, the riskier it gets. You never know what hidden flaws an image might have during training. But the number of images determines the variety of ways something can be drawn. I could go on philosophizing forever, but I’ll stop here.

By the way, I wanted to release a version for Pony—it does produce interesting results, but Pony probably needs a much larger dataset than Illustrious, so I’ll hold off until the dataset is truly impressive.

I’d love to have some communication with those who appreciate my work, and what better place than a convenient messenger? If people join, I’ll frequently share various materials there—maybe even datasets (if there’s enough interest).

Try to create small images - the result is great for me.

Version 6.3

I trained this version using datasets v4 and v5, but most importantly, I wanted to learn something new for myself. Instead of the usual 12 epochs and 1 repeat, I did the opposite: 1 epoch and 12 repeats. The results personally make me happy—it’s as if the generalization of epochs didn’t improve things, and the final output doesn’t look like a clone of the dataset, even though there were only 723 images in it.

Just in case, I recommend using my parameters for generation: Euler_a, simple, checkpoint: PlantMilkSuite_walnut. The number of steps is optional, but usually more = better, although 26 is often enough for me. CFG may depend on the model, but for my favorite model, CFG 4 is best. If you want to add your own LoRA, think twice—was it trained with any kind of smoothing? Pixels might just vanish under the pressure of blurring. It’s probably best to avoid that LoRA if it kills pixels.

Also, I should note that even though you can activate LoRA with any checkpoint, most of the ones I’ve come across don’t work well with my model. Keep that in mind.

Unfortunately for those who use artist tags in their prompts, those tags will likely blur your image too. The best solution is to avoid artist tags altogether. My recommendation for prompts: start with "pixpix, 8-bit, pixel_art" and end with "masterprice"—this way, the image stays sharp, and the pixels won’t die.

Now, onto a more sensitive topic for me. My model seems to be growing like yeast, judging by CivitAI stats. I have plenty of ideas for implementation and further development, but there are a lot of nuances I’ll get into now.

First, I’m going crazy from the lack of speed, so if any of you want to help with your computing power, message me—I respond to everyone.

Second, I’ve tested many base models for training. Trust me, Illustrious is not the best model for pixel art. I’ve tried SDXL, PONY, and Illustrious, and here’s what I found:

Pony is a fascinating case—maybe its understanding of the world isn’t perfect, but its artistic output is impressive, and its biggest strength is how well it absorbs materials. Pony is the closest any model has come to the "GameDev" space, which is crucial. Of course, Pony is far behind Illustrious when it comes to NSFW content. Illustrious knows anime and characters well but struggles with backgrounds compared to Pony.
As for SDXL? I don’t even know what to take from it. It turned out to be too complex for me, so I honestly don’t know what to do with it. I might just upload it for fun—you can play around with it yourselves.

By the way, I’d be incredibly grateful if any of you published your work using my LoRAs. It’s important to me—this way, I can see what we’re achieving together. It’s one thing when I get good results, but it’s another when you should too. Though, in half the cases, the generations don’t turn out so well. Maybe it’s because I only publish the top 10% of what I generate? Who knows.

I didn’t have a specific goal in mind for this training, like "training for game graphics." I just wanted to throw together a mix of images and generalize the essence of pixel art. But this topic has gone much further than I expected. My understanding of the situation could probably solve the entire problem of AI-generated pixel art, but as I’ve said before, I lack the hardware to pull me out of despair.

My current models exist only because I once got lucky and bought about 20,000 buzz from a friend. That’s why I’m training all my models on CivitAI right now, but even the smallest training costs 500 buzz, which is quite expensive. And there are so many things I want to test.

The saddest part is that I can’t upload more than 1,000 images to the site for training. This could only be solved by switching to local training, so if anyone wants to help—message me. Maybe I can train through you.

Be sure to leave your comments under the model—I appreciate them, especially critical feedback on the current state of the model.

Someone left me a link to their custom node for ComfyUI. I tried their tool, and while it’s more resource-intensive than what I use, it technically offers more features—just with a steeper setup. I think there’s a huge variety of nodes to choose from in this area, and you don’t have to limit yourself to what’s readily available. For example, I enjoy experimenting and sometimes use GPT chats to create my own tools, which is incredibly helpful, especially for sorting or automation tasks. The key is understanding the algorithm behind it.
https://github.com/HSDHCdev/ComfyUI-AI-Pixel-Art-Enhancer/tree/main

Below is a lot of information that could be useful—at least in theory—if you’re working with older versions. Just in case, I want to keep this as part of the record, even if it might feel like hoarding outdated material in some ways.

If my model isn’t working for you, there could be many reasons—whether it’s the model you’ve selected, an incorrectly configured setup in ComfyUI, or a broken Forge/Automatic installation. Before leaving a negative review, I recommend at least restarting your computer. You might not believe it, but that once helped me with VAE decode issues.

If you’re experiencing something specific and serious, don’t hesitate to mention it in the comments. Any problem can be solved. Additionally, your prompt and negative prompt can have a critical impact on the results.

You can even upload your output image directly under the model so I can see how disappointing your results are. Fortunately, CivitAI allows you to download the image and share your entire workflow from my end, which might be enough to diagnose the issue. I’m slightly obsessed—so I check my model’s page every day to see how things are going.

Version 5.0

I'm sorry, but I'm sincerely too lazy to live this life. Let me tell you, all the arts posted were made using the WD TAGGER node.

I have a lot to tell you. At the very least, I specifically made a custom node for our pixel needs, which perfectly helps me in ComfyUI with pixelization (of course, my node doesn't change the number of colors, but you can arrange this post-processing yourself if necessary).

I completely forgot to mention: use all three of these trigger words: pixpix, 8-bit, pixel_art.

Without them, your lora might not work at all.

I will leave the link to the node itself right here: https://civarchive.com/articles/17203

I'm sure its simplicity will make it indispensable.

I love the PlantMilkSuite_walnut model.

I can recommend the Euler_a sampler and the simple scheduler.

But here's what has really been boiling up...

Imagine that training pixel art using the "Lora" method is like working with a small part of the model. I don't change the model's understanding as it should be. Despite the fact that many of the results presented by me may seem interesting, the model itself knows what it knows, and the pixels in such training do not displace everything unnecessary. There are enough remnants of the model's primary knowledge for it to draw circles and lines in the eyes, trying to make the eyes look more like eyes than pixel art. In general, it's hard to describe, but the essence of my condition is as follows:

I have an incredibly large dataset that needs to be manually trimmed. But even finishing this task does not cancel the fact that I have always trained on CivitAI, and the site simply does not allow me to upload more than 1000 images. Technically, the only thing that could save me is training locally or on rented servers, but I am a wretch who still needs to be found, functioning solely on enthusiasm and a fixed idea. For your understanding, I have an AMD RX 6750 with 12 GB of VRAM. It takes me from one minute to two and a half minutes to create one image. I am trying to run local training even with such equipment, but I fear that only a computer upgrade could give us any results.

Honestly, as I delved into the topic of training, I found out a lot of interesting things. For example, there is a training method called fine-tuning. It is fundamentally unavailable to me, but I hope that when I get the hardware, I will be able to find those who could help me master such a powerful tool. If this happened, if I could retrain the entire model on all its blocks, layers, and touching every pixel, if I had such an incredibly powerful tool in my hands, I could create something greater and accessible to everyone.

I understand that such things probably should not be spoken aloud, but there is something called Retro Diffusion. I have nothing to say about it except that it has a plugin for Aseprite, a bunch of features and tricks that I can't even dream of. The only thing I wanted to say is that its services are colossally expensive ($65?!). Despite how much effort and nerves I spend, I would first like to help my past self. I went into neural networks literally only because I once wanted to make a game with pixel graphics. Now I am at the stage where I can train, guess which loopholes can be used even with the current limitations, and sincerely believe that in the future I will achieve better results than anyone else if I try and earn enough money for new hardware (with my own hump, sweat, and tears).

Just in case, I'll tell you, this version had only 257 images from completely new images in training, and this will greatly distinguish this version from others. In addition, Alpha and Dim were set to 64, which is why Lora should be strong by itself.
In addition to the above, I found a way for myself to make training extremely cheap within the framework of CivitAI, so this model was extremely important to me. All that was needed was to crop the images to no more than 1984.

Don't be afraid to experiment; I sometimes test random LoRA models at a low influence level. Values like 0.1 to 0.3 shouldn't disrupt the pixel image too much, but it can add the vibe you need to the picture.

By the way, I was told that version 4 had well-functioning tags related to artists' names. I apologize if you have issues with this in version 5 and newer. I've never even thought about such features, and technically, I don't fully understand how to use them.

I sympathize with Forge and Automatic users, as I have nothing to say to you. I can't use these interfaces, so there will be no recommendations for you.

Sorry about the messy description - too tired to clean it up. Old notes below might help if you're still using older versions.
------------------------------------------------------------------------------------------------------

I strongly recommend trying generation with the 8-bit token. In my mind, it seemed like something was off. In reality, it actually helps with generation. However, the question here is more about your goal and whether you need such graphics.

I am pleased with the results at these values:
euler_a

sgm_uniform

step: 36

Model: PlantMilkSuite_walnut OR WAI-NSFW-illustrious-SDXL

To my surprise, the model can generate pixels even without words related to pixels. It works like this: the word "pixel_art" excites the model's desire to draw in the way it knows pixel art, but the thing is, I didn't teach the model pixel art. I only specified the trigger word "pixpix" in the training, and personally, I don't need to write it for the Lora to truly work. It is enough that Lora is enabled.

Words like "Pixel_art" and similar ones can still make the neural network's job easier because contours and limitations that pixel art is prone to appear.

I apologize if I am asking too much of you. Please publish as many of your works as possible under my model. Leave comments, tell me what you don't like, and what you would like to see. I listen to all your words and will be happy with any of your works. I love you for your activity. Thanks to it, I have more opportunities to test.

I really wanted to give you a universal tool, and without your support, it will take too much effort and time to build buzz.

Next on the plan: tests, tests, and more tests. Various parameters can potentially be useful for training. I haven't even tried describing what's happening in the training materials yet. Well, well, well. I don't even know if it's necessary at all.

Version 4.0

This version is notable for its very wide range of data, comprising 488 images. Its current state is capable of demonstrating valuable results. However, I was interested in looking at this version from a different angle. Instead of the old policy where I gradually increased the dim and alpha parameters, in this version, I changed them to alpha 32 x dim 48. Is this sufficient for implementing the model? It's difficult for me to answer this question because I don't have specific real tasks required by someone in particular. Therefore, I am not limiting myself or you in this LORA model.

Version 3.1

It was important for me to make sure of some things, hence this version appeared. It is unique in that the dataset is completely different from previous versions. The styles are actually different, which is very important. Although, to be honest, I'm not sure why the pixelation misbehaves in half of the cases and does not reach the level I need (and I thought I would achieve the result I needed, but it seems not this time). Only time will tell how much sense there is in my actions. Please, evaluate with me this essentially completely new (but test) style. By the way, it makes sense to use all the trigger words that I specified. In any case, I have good results.
Why did I decide to name it 3.1? It was created less seriously than the others.

Version 3

I corrected the dataset as best as I could; working with 280 images is not that easy (there were enough mistakes). The goal of this endeavor was to expand the potential of the model. The more faces you feed the model, the less likely it is to draw the same thing over and over. Then I decided that it would be a good idea to add some more works; later, I encountered a significant difference. The second version of Lora, it turns out, includes not only pixelation but also a limited palette. Version 3 differs not only in this regard; in some places, it is brighter and richer. In any case, the visual style has changed; decide for yourself which you like more.

I tried to slow down the training, changed the noise offset, increased the dim, but was afraid to touch the alpha. I am considering various possible solutions to the problem of why pixels do not want to be pixelated and get blurred. My biggest fear is the words of my mentor, who believes that the neural network does not remember more than one megapixel of information from a picture. He might be right. If so, we will have to train on works smaller than 1024x1024, which, if multiplied, gives 1,048,576, and 1 megapixel is 1,000,000 pixels. We will have to work hard to adequately train the neural network in such a narrow format. I do not have many sources of high-quality pixel art.

Unlike many other pixel models you can find on CivitAI, mine can boast clear rules for adhering to pixel art, but unfortunately, I have not yet had the strength to tackle testing the control of "Dithering"; for now, it will appear spontaneously. I am sure it can ideally be controlled just like the presence or absence of outlines; for now, the model prefers outlines.

Version 2

I expanded the dataset to 281 images (there were about 50 in the first version). The variety of possible generations has increased significantly, but the pixels are still not perfect. And you know what? I intend to review absolutely all the images in the dataset because, in my opinion, the problem might lie in hidden imperfections. If even one of the images has a slight gradient or a broken pixel, the entire training process could be ruined. But what terrifies me even more is that I can't find a similar "pixelate+" tool. For example, in paint.net, there is a "pixelate+" effect, and the pixelation method there is so good that the image is perfectly pixelated almost without changing. And I'm very sorry that I haven't been able to find a similar tool in Comfy yet. If you use the RESIZE method like I do, I have bad news for you—this method is not ideal. At least, that's how it was in my tests. Our conditions might differ, and in reality, everything might work out great for you.

Version 1 (test)

Frankly, it turned out not quite what I want to achieve in the next version.

You can wait for the release of version 2, it will be more worthy than the current one.

If you are not scared by the words above, then know that the test LORA is strange, over-trained and tends to draw much more often what it was trained on than to match text prompts, keep this in mind.

You will encounter many swords, strange books, Easter Island statues during exploitation and, most importantly: GigaChad (even a girl can acquire his facial features, but this is uncontrollable (in any case, I didn't try)).

Remember that each version is unique in its own way. Each had its own dataset, and some handle certain themes better, while others excel with different topics. Some versions work more easily with various checkpoints, and let’s not even mention LoRAs, which also interact differently with my style every time. Don’t hesitate to try everything.

Contact with me:

https://t.me/SKORMINO

Files with SHA256: 9384de887cd6164fcbc0f6495599e7cb80ee88c890228dbbdf72d3d60095e3e0