現状、pytorch-CycleGAN-and-pix2pixも万能ではなく、SDXL+自作ControlNetの方が優秀ととれる場面もあった
一枚目が自作ControlNet、二枚目がpytorch-CycleGAN-and-pix2pix
これは線画から疑似3D画像に変換するタスク

1 0

CycleGANってその仕組上、A→Bの変換器とB→Aの変換器も作るわけで。今回、生身→石化の変換しようとしたら、石化→生身の変換もできる。やってみるとこんな感じ。
石のテクスチャ消して肌を作ろうとしてるの、すごくない?(髪の色もその部分は変えようとしてる)

18 54

CycleGANで学習用画像のサイズを256x256に合わせる。
んだけど。正方形に変更するのに画像のアスペクト比を保っておきたい。とすると左右(上下)に黒帯を入れる(1枚目)、というのが簡単だけど。これやると特徴が変わりそう。
なので、同じ画像を並べてみよう、とやってみた(2枚目)

1 2

Hago fotos y dibujos tipo cuadro con inteligencia artificial, a base de oscilaciones (Mezcla de imágenes de un database y una alternativa modificada de cyclegan a la inversa + photoshop y capture one)
menciono a

6 6

AdaIN-Switchable CycleGAN for Efficient Unsupervised Low-Dose CT Denoising. https://t.co/yfMO0wcT6t

4 4

i ran 3,000 panels each of CAD and Garfield through ganilla's cyclegan implementation, hoping to transfer each comic's style to the other

disappointingly, the results for cad2garfield were: desaturated colors, an attempt at shading, and gibberish WORDSWORDSWORDSWORDS

1 10

Warning: flashy things/strobing

This is 100 style samples from MUNIT (a multimodal version of something like CycleGAN). Lots of them are garbage but its really nice to have the selection!

0 6

Comparing overnight cycleGAN training before my morning coffee, can you tell my priorities?
(1) Digital collage generated with code
(2) GAN Output trained on 20th century still life paintings (Matisse, Cezanne)
(3) GAN Output trained on 48 of my own still life paintings

0 13

soulful tagine...
(that's from when i trained cycleGAN to draw faces on my stills)

1 12

might be a magic formula:
SNGAN_projection+CycleGAN=pen&inkGAN

1 8

A second series of AI generated nude portraits 🥩🎥

It's painting with warmer tones and better colours (imo), but still hasn't mastered the human form.

Working on other neural networks to post process these- I have some neat stuff planned with cycleGAN that should be done soon.

81 287