GENERATIVE MEANS

2. It seems to be Generative "encoder-decoder" technology works on Plato's philosophical principles of idealism. Otherwise it hard to explain. It extracts idea of objects then generate new objects on the basis of this idea.

3. How are deep fakes work? It works on GAN basis. But we put together (generator trained on zebras) and (descriminator trained on horses) so we get zebra to horse "translation". Use CycleGAN technology.

4. What generative models i know. Stable Diffusion Stability AI, DALL-E2, Imagen, Parti, Muse, LDM, DreamBooth (text to image), GPT, PaLM, Gopher, BART, T5-XXL, MT-NLG, Luminous, BLOOM, Jurasic-1, LaMDA, OPT-175B BB3, GLM-130B, GAL, YaLM, NLLB, Flamingo, Chinchilla (prompt to text), Minerva (solve math), Phenaki (text to video), DreamFusion (text to 3D), Super-Resolution (low res image to high resolution one)

IMAGES

Diffusers

1. Diffusion creates random image from random noise.

2. To create image on base of test prompt diffuser model uses zero-shot classification model like CLIP. It is on each step generates couple of image and calculates loss with zero-shot classifier then calculate which image is closer to the text prompt.

3. With diffusers you can generate also video and even audio (spectogram).

Keras CV Stable Diffusion

0. Web API in stability.ai

0.1. SD uses VAE principle.

0.2. You can use maximum 77 words text prompt.

0.3. Could generate image from text and from another image and text, could change only masked area of the image.

1. To install $ pip install keras_cv

1.1. Or with HF-hub Transformers API: "stabilityai/stable-diffusion-2"

2. While I am trying to create model = keras_cv.models.StableDiffusion(img_width=512, img_height=512) I get an error

qt.qpa.plugin: Could not load the Qt platform plugin "xcb" in "/home/john/django/lib/python3.8/site-packages/cv2/qt/plugins" even though it was found.

This application failed to start because no Qt platform plugin could be initialized. Reinstalling the application may fix this problem.

Available platform plugins are: xcb, eglfs, linuxfb, minimal, minimalegl, offscreen, vnc, wayland-egl, wayland, wayland-xcomposite-egl, wayland-xcomposite-glx, webgl.

Compatibility problem, use new virtual enviroment

3. images = model.text_to_image("photograph engineer", batch_size=3) takes too much resources. Try to decrease image size but it should be multiple to 128 (128, 256, 512). // helped

3.1. Add jit_compile=True, m = keras_cv.models.StableDiffusion(img_width=128, img_height=128,jit_compile=True) // Didn't help and lead to error

4. Try mixed precision keras.mixed_precision.set_global_policy("mixed_float16") // Lead to crash

5. To create image with specific face on diffusion type of model use Dreambooth.

eDiffi

1. Allows to generate image from draft image with words.

DiT

google/ddpm-celebahq-256

1. Generate faces.

2. Usage

image_pipe = DDPMPipeline.from_pretrained("google/ddpm-celebahq-256")

images = image_pipe().images

images[0]

DALL-E

Google Imagen

NEURAL STYLE TRANSFER

1. Use pretrained model kaggle.com/models/google/arbitrary-image-stylization-v1/tensorFlow1/256/2

2. DeepDream is similar technic but has something common with fractals.