-
-
Notifications
You must be signed in to change notification settings - Fork 408
Glossary
JetVarimax edited this page Jul 25, 2023
·
3 revisions
-
Inpainting: latent text-to-image diffusion model
Select part of image and replace it with semantically generated context based on output prompt - Outpainting: technique to increase canvas and then use inpainting to fill missing parts
- Upscale: run resulting image through additional super-size ML model to increase resolution
-
Textual inversion: learn to generate specific concepts (objects, styles, persons)
by describing them using new words in the embedding space of a pre-trained model
creates embeddings assigned to one or more tokens from sample images - Diffusers: used to synthesize results by applying series of applications of denoising autoencoders
-
Latent Diffusers: its basically using diffusers in latent (abstract space) before generating pixel space
simply more efficient than running diffusers in pixel space - Conditioning or Encoding: text or image to semantic map
- Transformers: generic ML model that add semantic understanding to trained area (text or image or audio or whatever)
-
Checkpoint": when training a model, save it as checkpoint every n epochs so training can be continued from there
checkpoint models can be further trained or used as-is -
Finetune model: adds specific retraining using sample images to existing model
different than full retraining as it starts with existing checkpoint -
Hypernetwork: finetune model and save as extension model instead of modifying original
this is basically an adaptive head - it takes information from late in the model but injects information from the prompt 'skipping' the rest of the model
similar to fine tuning the last 2 layers of a model but it gets much more signal from the prompt -
Dreambooth: essentially model fine tuning, which changes the weights of the main model
differs from typical fine tuning in that in tries to keep from forgetting/overwriting adjacent concepts during the tuning -
Sampler: which algorithm or lightweight ML model to use to add noise in each step before diffusion
different samplers are better at specific steps ranges and styles