AI lessons

I used to want to train LoRa on a computer. I tried it on Kohya, but there were errors. I saw various tutorials on training LoRa on paid websites when I was taking various AI courses, but I wanted to train it on a computer.
I decided to try training LoRa in ComfyUI.
In Cloud, training is on runninghub.ai it doesn't work, there aren't many nodes.
I decided to create a LoRa model of Petryk Pyatochkin from the cartoon.
How Petya Pyatochkin considered an elephant (1984) Cartoon by Alexander Viken. Original video
How Petryk Pyatochkin counted elephants (1984) Original video
How Petryk Pyatochkin Counted Elephants in the UKR version FullHD / How Petryk Pyatochkin Counted Elephants
The cartoon is low quality; all the images are very blurry and weak for LoRa training.
First, I prepared 30 images up to 800 pixels in size. I tried to simply take screenshots of Petrik Pyatochkin's images from different angles and generate LoRa. I tried different image sizes and tweaked various settings, but for some reason, LoRa generation on my graphics card took a very long time, taking two days to complete. I started and stopped LoRa generation several times. Then I decided to set the minimum size to 512x512, select 20 high-quality images of Petrik Pyatochkin, and refine them before training LoRa. I used AI upscaling, improving the graphics of the images for LoRa training.
One LoRa model was trained 28 hours to train on my RTX 3060 12 GB graphics card, which is quite a long time, from the author's model, with higher settings and a 16 GB GeForce 5050, trained for 28 minutes.
I managed to train Petrik Pyatochkin's LoRa model, but it only works on flux1-dev.safetensors, which is what it was trained on. It can be run on compressed versions of flux1. I created the LoRa model on flux1, so this model won't run on other models or on flux2. To train Petrik Pyatochkin's LoRa model, I tried different angles, upscaled the images, and also used some images not from the cartoon.
In different images used for training LoRa, Petrik Pyatochkin looks different. His torso below the waist is very different: sometimes it resembles a triangular shape, sometimes it has a more realistic, proportional appearance.
A noticeable distinguishing feature of all the characters in the cartoon is the unique shape of the chin, the oval of the face, and the neck. All the characters have an unusual chin, which resembles a triangle, and their heads are located on a thin neck.
But the AI ​​doesn't take this into account when generating images. When I was preparing images for Lora training, some images lost the chin outline and leg shape when the graphics were enhanced. It's clear that in some images, Petrik Pyatochkin has very short legs, with virtually no knees. In other images, knees appear, and in some images, he has long arms and legs, like those of ordinary boys. Therefore, the trained Lora often generated a more realistic torso shape with legs for a boy. However, there are cases with short legs and three legs, with one added somewhere on the side or back, as in the images in the cartoon where Petrik Pyatochkin runs with altered body proportions. Also, Petrik's oval face shape isn't always chinless; it's often rounded, like that of a realistic boy.
Now you can use Petrik Pyatochkin's image on various models, for example, on the model:
AI Image Editing in ComfyUI: Flux 2 Klein (Ep04) or on qwen, z-image, you can change the character's poses without LoRa model.
The LoRa model can be trained not only on one character, but on the entire cartoon style, as the author did.
She created the first Lora model for the entire cartoon style. "Treasure Island" (1988); the author trained the second Lora model on a single character.