( Not my Model )

https://civitai.com/models/1456548?modelVersionId=2468132

• Dixar version 3 is a trained model ...mostly. It is a merge of five separate trained versions of Dixar.

Each version was trained on Kohya. I am still learning how to train checkpoints and honestly, I kinda suck at it. I have trained it about 20 separate times and I have six versions that I like but none are perfect as standalone models. This is probably among the best I have so far. The dataset at the time consisted of 2600 images, 90% of the dataset was created by me, which is where I think the model is unique, very high quality images in a dataset, almost all of which were created by me. They were created using everything you can imagine, some in Comfy with various versions of Dixar that I have, some using Chat GPT, Some using Nano Banana Pro, Hunyuan image 3, Wan2.5 image, Seed Dream 4, Qwen, and Midjourney.

The 10% of the images in the dataset that were not created by me are screenshots from 3D movies. However, they are not the screenshots you would find online. All of the screenshots in the dataset were taken by me because I didn’t like the quality you can find online and I have a true 4K monitor. Every screenshot was also hand edited by me. And yes, I had to watch every 3D movie to take the screenshots.

Just building the dataset probably cost a few thousand $$$ worth of gpu time or credits on various sites, as well as a ton of time. Currently I have brought the dataset up to 6910 images but I have not trained, or even organized those images yet.

This version defiantly has some issues, but it is good enough to post.

It aims to be closer to a screenshot from a Pixar movie, it's obviously not there yet but it is getting there.

You don't have to specify 3d, Pixar, or anything but I feel it looks better when you do, just like anything with image generation you are telling the model what you want it to make for you.

Euler A

Beta or SGM

50 Steps

3 CFG

It does not know specific characters, although I only tried one.

It is also fragile and easily breakable. It will work with some loras but others just break it.

If you are seeing this within a few days of me posting the checkpoint, like half the example images got flagged for review.

The very last image is a raw batch of 50 images.

Just credit me if you merge it.

I Like to see the images you create, plus if I like them a lot, I'll add them to my master dataset for v4.

As always, have a wonderful day and go fuck yourself.