#Hey im new and I have two questions

10 messages · Page 1 of 1 (latest)

shell sleet
#
  1. Would it be possible to train the AI with an custom data set if i provide for example 10k images with tags.
  2. Is it possible to do this with an RX 6700 XT (AMD GPU). I am new to AI image generation and im currently just researching about this
supple phoenix
#

👋 I don't know too much about it, but here's what I think:

  1. Other model variations were finetuned on sets of 30-40k images. So 10k might work, but maybe not super-awesome. :3 Not sure.

  2. That sounds harder.

  • There is no easy, documented way to train the model with invokeai. If you train a model, you can use it with invokeai though.
  • AMD cards are not supported yet (right?), there's only testing for it. So that might be hard.
  • I think training the model takes a loooot of hardware power and vram, so I doubt it'll work on normal consumer hardware. :X

There is, however:

A) Embeddings. You can find those in the documentation for invokeai. Basically: Give the ai "5 images + 1 tag + 1h training on a RTX 3090ti = ai learned new word".

B) Dreambooth. I don't understand it yet, but it's kinda "embeddings but better", if I understood it correctly. It is not implemented yet, but it will be implemented.

nova harbor
supple phoenix
dusk vault
#

If you want to train it on many different images I would use this instead of Textual Inversion

cinder bronze
#
  1. FYI. After two days of research I can successfully run InvokeAI on Radeon 6700XT on Manjaro (arch-based).
  1. always use export HSA_OVERRIDE_GFX_VERSION=10.3.0; preambule.

As I found out the instruction set of 6700XT is same as 6800XT's and the latter have official ROCm support. So this trick will always work.

  1. install ROCm
  2. install InvokeAI
  3. upgrade pytorch to the ROCm version in conda.

EDIT: ~15 sec per image (512, defaults)

#

@shell sleet FYI

nova harbor