Embeddings

[JohnTigue]

Embeddings seem to be like finetuned models, but different. Some folks seems to be expressing preferences for embeddings over models. No idea what is going on.

• Textual inversion: training an embedding
• Ultimate Free Textual Inversion In Stable Diffusion! Your Face Inside All Models! (YouTube, 25min)
• Detailed guide on training embeddings on a person's likeness

[JohnTigue]

Looks like embeddings are just another name for models trained via textual inversion. Might just have a dup issue here, see #2.

[JohnTigue]

But embeddings are small files (say, ~10KB) which is much smaller than full models (gigabytes). They can be either style or object centric.

[JohnTigue]

Embeddings seem to be a SD2.1 thing, or at least they work together well. The embedding files are very small (tens of kilobytes) compared to fine tuned full models (~4GB). The have trigger words (often of the form foo-style) which can clearly demonstrate what they do. Powerful. Kinda like mix-ins.

[JohnTigue]

Seems both embeddings and checkpoints are still being trained by textual inversion AND Dreambooth, given that InvokeAI 2.2 is planning on support both.

[JohnTigue]

In the finetuning space, it seems textual inversion has recently pulled up alongside DreamBooth which was supposedly gave better results in the SD 1.x timeframe.

Seems that textual inversion embedding took off with SD2.0. In earlier SD 1.x releases, the results were not as impressive as using DreamBooth. But once SD2.0, they were competitive with DreamBooth. Additionally, DreamBooth gives you a whole new multi-gigabyte model checkpoint file (.ckpt), versus embedding which are on the order of kilobytes [*].