AI, human dance generation. Maya's dances, 20 songs, Summer 2026

This is my little niece Maya. In the photos, she is only 1 hour and 5 months old.
In my opinion, it is very well suited for generating ideas, tested clothes, moved.
She can be suitable for both male and female children's images, be in different clothes, perform children's and teenage skits, as well as adult dances. I generated dances with my niece, I looked for different beautiful popular melodies of any genre, in any language, the main thing was that it was melodious, there was an effective dance, it was suitable for a child and it was interesting to watch and listen.
The most important thing in these videos for me was to make not just movements, but something interesting in the movements, to come up with scenes to keep attention, to make them realistic, as if a child is dancing and has reached an extraordinary level of mastery. Even when I watched the generated dances with Maya, I pressed pauses and tried to understand how dancers can control their bodies in such a way, because it is necessary to constantly think about the movements of different parts of the body and do it beautifully, how much effort and training is invested in a professional dance.
The purpose of these generations was to test the generation of graphics and video, to come up with an idea for the dance, the image of the girl and the atmosphere of the scene, to generate parts of the video and edit interesting parts from the video in order to better understand where failures occur and what difficulties there may be, to analyze how the camera behaves and what is the limit of its capabilities on the model I use, what failures occur and how to get rid of them through prompts or changes to the example from the video itself.

It takes a long time to generate a video on free neurons on my computer. If there was a very powerful video card, it would be possible to download the largest Scail2 model with a volume of 32 gigabytes and try, but complex scenes on my video card cannot be done in good quality on free models. When generating images of faces and video, it is not immediately possible for the faces to match well, you need to test a lot, sort through, look for a good option. The Flux2Klein generation replaces faces quite well, you can choose options, when generating a video in one video the face may fit, in another video the face may not always match, the proportion of the head may change, the head may be flattened, the eyes may shrink for some reason, the mouth may open arbitrarily in the dance, the smile may disappear, different mouth movements may be added, the person may look away, lower his gaze. Sometimes it turns out successfully, sometimes not. Although in the example the girl can look at the camera, in the generated version the girl can do something wrong, it is necessary to clarify the failures in the prompts and generate anew, watch the preliminary generation and continue to look for a successful animation. Also, if the video is long, for example, I generated some variants for 20 seconds, then at the end the contrast of the entire video changed, the face changed its color, became darker. Sometimes the face of the girl in the dance was clear, and sometimes it was not possible to achieve clarity, the face changed. I generated some videos in parts of 5 seconds each so as not to wait long. But in this case, the following parts of the video no longer matched the position of the body, different details in the animation changed, and the face differed. These changes were minor, but they were, for example, the hairstyle was changed where it was poorly visible, various small details in the clothes were changed because it was already a new seed.
Better to generate video on paid sites: Kling 3, Suno Minimax, Seedane 2 and others.
It is possible to generate graphics on flux2klein9, quite correct results are obtained.
If you do not need to generate a lot, then you can generate on a computer, if you need a lot of generation with very high-quality stable results, it is better to use paid sites: Nano Banana, Gpt image 2.0 and others.
For audio generation, you can use Suno, Eleven Labs, Minimax Audio. You can use Claude, ChatGpt, Gemini, Kimi to generate prompts.

There are paid models for generating graphics and video, but I used all available free models that I knew, since these are not commercial developments.
At the time of making these videos, my niece was one and five months old.
At first I wanted to make an entertaining video with a girl dancing, but then when I started having different ideas, they accumulated so much that I wanted to make them all.
New and new ideas for dances appeared.
I decided to make them professionally, as if I were creating a mini-album with girls' dances, and then I would add all of them to my portfolio. I wanted to stop the moment for my little niece Maya and fix her image in song dances, how many ideas can be implemented. Perhaps she will like some video later.

I generated graphics on flux2klein9, used claude.ai and chatgpt.com to generate prompts and Scail2 model from the lesson
ComfyUI Scail 2 Video: WiFi Nodes, New Groups & Loops (Ep23)

I had 4 photos of the girl, based on which I generated all the images and videos, tested them, tried to achieve similarities and correct movements in the dance and in other body movements. In some videos where the image was generated and not edited, a similar face was also obtained in the video, but not always, in different images it turns out differently. But it happened that I edited the face in Photoshop and it was difficult to animate it, it constantly changed a lot and was not similar, it was completely different. For some reason, some faces in the images almost always look similar in the video, while others are very different and need to be generated and sorted a lot to achieve similarity.

After studying many lessons on graphics and video generation, I decided to test the creation of human dances based on examples. I was looking for different short clips with people dancing to well-known popular songs and wanted to practice what could be done and what was difficult to implement. Then I wanted to make not just a short video of one dance, but real clips, mix different scenes, combine different angles. Vertical and horizontal 512x896 videos can be made on Scail2, my video card cannot handle larger sizes. You can upscale the generated video. Paid models generate better and faster. I used the best free models in ComfyUI that were available at the time of creating the video. At first, when I was generating a dance, there were videos where sometimes people danced, parts of their bodies would stick out behind the screen, the camera could shoot a person too close or far away, move to the sides and not keep the person in the center or anywhere with the dimensions I needed. Then it occurred to me that you can simply create a video in Premiere Pro with the correct location of the person in the frame, adjust the entire animation to the correct position in seconds, zoom, move. In this way, it is possible to generate a good dance from a filmed dance with beautiful movements, where the person will always be in focus and zoom in, remove the camera as you need, place the person as you want at any moment of the video, and even adapt the dance to vertical or horizontal size. If some minor body parts are not included in the video, they will still be automatically added. Video generation will repeat all this on the modified video.

Often, when generating video on Scail2, the camera is fixed in relation to the background, for example, a curtain or a view of a park with trees. In such videos, the camera did not float. But in one video where there was a river and a boat, I generated 3 parts of the video, since the entire video was not generated at once. The camera constantly moved to the right and away from the sun, directed towards the flow of the river, although I wrote in the prompt that the camera does not move. It turned out that each part started anew from the place where the camera should be, and the camera at the end of each part turned out to be displaced. I did not test this scene for a long time, maybe there was a boat, there was a river, for some reason the AI ​​constantly brought the camera into that scene, it was not possible to keep the sunset with the sun in focus. More time was needed for tests, but generation is slow. In the second video, there was no such displacement of the camera. But in the video with the boat, an interesting result was obtained, which can also be used, the camera rotates a little around the person for no reason.
Also, if you start the video for 8-10 seconds, it starts to deteriorate at the end, more and more inaccuracies appear, the video becomes darker, artifacts appear on the objects. In order for this not to happen, seconds need to be shortened. Sometimes it generated normally, sometimes not. But if you cut the video into parts, then each generation changes the person's face and it may not match, it may slightly shift outside of the person. Therefore, on the one hand, I want to upload one long video with an example of movement and generate it at once, but on the other hand, it is generated for a long time, there is no way to check generation errors during generation, only the first two seconds can be checked. On the one hand, it is faster to generate a 3-5 second video, but on the other hand, if there is a 15-30 second video with an example of movement, then you want to generate it at once, but the generation will take a lot of time and the entire video may have various flaws that cannot be corrected separately, while small pieces of video are easier to check and regenerate only those where there are failures.