HighCWu 艾梦 an Amateurish Researcherさんのイラストまとめ

Add residual to latent diffusion first stage training and train with danbooru dataset. Now it can encode 256x256x3 pictures to eight ordered 16x16x4 latent codes. I don't know if this first stage model generated latent codes can be used in diffusion. https://t.co/5NFcO64qDK

Prompts: Beautiful anime girl
It looks like the (left) is not as good as the XL version. However, I saw that some of the generated results (right) without self-reranking are better than those after reranking. It should be just that the self-reranking ability is not good enough.

Tired of Colab K80. So I create a repo of PaddlePaddle version of ruDalle. The weights were converted in order to use the free v100 inference.
Github: https://t.co/jW9iYGOTpf
Free v100 on AI Studio: https://t.co/Lc48Kr9mPW (Warning: This is a Chinese community)

Recently, I saw some papers on the separation of image colors and sketches, so I sorted out an idea that I had two years ago, and opened it up: https://t.co/vX3jgtZZ1d
It's a neural network model for extracting regions from an image. The size of this simple model is just 600KB.

@ak92501 It seems to coincide with my idea, but mine is the coloring of any anime line drafts. The effect of this article looks a bit poor. Encoder of the line draft should not be needed. Just merge the line draft and noise together as input.
https://t.co/mOJWWIeANK

For very short ddim steps , my new anime colorization model sometimes generates images with noise.

I thought the colorized results were not good on real datas, but I was wrong. Today I found some real anime-sketch pairs datas for comparison. The colorized results of this model is truly amazing to me.
Results in https://t.co/2n9I8EQcrp with `sample-chaosinism*` prefix

A comparison with style2paints, although my model does not use hint points. In the middle are my model generated, it seems that my model does not work as I imagined.

I tried to use the denoising diffusion model to color the animation line art, and it looks good (it would be better if I cleaned up the data). Use 128px danbooru2017 and the line sketches converted using sketchkeras as the training set. (Second rows are colorized images)