ControlNet VS lora

It`s all possible. Have a look into Automatic1111`s Web UI, ControlNet, OpenPose and, if you don`t have a dedicated GPU with at least 8GB of VRAM, or at least 16GB of RAM to use the CPU, you can also use Stable Horde to use the webUI with a peer-to-peer connection, where you`ll only use a fraction of your resources, but you`ll be able to use local AI models with all the bells and whistles that you won`t get from "state-of-the-art" paid services.

Next up is ControlNet. Controlnet, as Illyasviel--creator of controlnet--describes it, "let's us control diffusion models!." ControlNet is a neural network structure to control diffusion models by adding extra connections. [8]. There is more to that than what I described, but the big take-away is that ControlNet takes a preprocessed image that you provide (or is generated) and uses that as a way of constraining the output the sampler's noise generates, allowing you to have a bit more control of the output. ControlNet is typically used for character or scene "artwork", which previously would have been a challenge with just prompting alone (at least with this current architecture).

For your first part of the comment, I can simply refer you to technologies like ControlNet, LoRA and prompt embedding: https://github.com/lllyasviel/ControlNet https://github.com/microsoft/LoRA

Here you can see what every controlnet preprocessor and model do, to give you an idea of ​​how to use

Ya, it takes awhile to really start getting comfortable with the wonkiness. If you are trying to do something specific, look for a LoRA, but in general I'd recommend you get controlnet so you can feed it a reference image. Another simple trick is to edit the image a bit in GIMP or a photo editor to get the color scheme you like and then feed it back to img2img at low denoising (0.1-0.2) to refine it. You can also add just garishly bad cartoon drawing or photoshop in assets and img2img will usually make something of them and blend them into your image, I find this easier than using img2img scribble.

If there are no errors, go to the "Extensions" tab, then "Install from URL". There, enter "https://github.com/lllyasviel/ControlNet" then click "Install".

This is a good place to start reading. Given the open-source nature of SD, there are setups of various difficulty available. A1111 is the "standard" people enjoy because it's easy to plug in new stuff (ControlNet, new models, etc.), but it's not inherently easy to set up and get going. There is an installer for it, but I haven't tried it.

Diffusion unet has an "extended" version nowadays that applies to the resnet part as well as the cross-attention: https://github.com/cloneofsimo/lora

Absolutely. But that doesn't matter because you only have to train it at scale, once. There are papers released already that show it's possible to update weights in small sections. You won't have to wait for the next monolithic LLM to drop to get up to date information. It will start to learn in bits and pieces.

No, it's a deep learning thing

Low Rank Adaptation of Large Language Models.

Recently, I have seen the LoRA technique (Low-Rank Adaptation of Large Language Models) as a popular method for fine-tuning LLMs and other models.

Enter LoRA, a method proposed for adapting pre-trained models to specific tasks[2]. By freezing pre-trained model weights and injecting trainable rank decomposition matrices into the transformer architecture, LoRA can reduce the number of trainable parameters and the GPU memory requirement, making the adaptation of LLMs for downstream tasks more feasible.

Open-source LLM projects have largely solved this using Low-Rank Adaptation of Large Language Models (LoRA): https://arxiv.org/abs/2106.09685

Apparently an RTX 4090 running overnight is sufficient to produce a fine-tuned model that can spit out new Harry Potter stories, or whatever...

Alpaca was trained for $600 ($100 for the smaller model) and offers outputs competitive with ChatGTP. https://arxiv.org/abs/2106.09685