Jesse Andrews
Jesse Andrews
When you train with SD2.1 (not SD2.1 base), the loss instantly `NaN` ``` Steps: 0%| | 0/3000 [00:00
this is a POC for testing what changes are needed to train for multiple concepts If your `instance_data` has filenames: "foo bar.jpg" -> "foo bar" "foo_bar.jpg" -> "foo bar" "foo...
When running with ckpt trained by someone else, it appears it can OOM while generating the default 50 class images. ``` Running predict()... Generating class images: 0%| | 0/13 [00:00
if training data includes a HEIC files, it will fail: ``` Traceback (most recent call last): File "/root/.pyenv/versions/3.10.9/lib/python3.10/site-packages/cog/server/worker.py", line 209, in _predict result = self._predictor.predict(**payload) File "/src/predictor.py", line 289, in...
NOTE: we are currently NOT using `COG_WEIGHTS` as we either double download or are stuck with a slow download This template model uses weights specified by URL set in`DREAMBOOTH_WEIGHTS`
Users have asked for support for longer prompts This might be the answer https://github.com/huggingface/diffusers/tree/main/examples/community#stable-diffusion-mega
To fully customize prompts, there is a "hack" that popular versions of stable diffusion use to allow you to increase/decrease the strength of each token using parens/square brakets [lpw_stable_diffusion.py](https://github.com/huggingface/diffusers/blob/main/examples/community/lpw_stable_diffusion.py) is...
It would be nice if you could mask/inpaint with your dreambooths. while stable diffusion 1.5 / 2.x has moved to having specialized models/weights for inpainting, as @andreasjansson says: > yes,...
ensure template works for SD2.1 that has Float32 weights. Early feedback from users is that SD2.1 training is superior with float32
When I use replicate.com I can see predictions for a specific version: https://replicate.com/predictions?version=7a5d96bdd41ffe8c449f6c6818a9aaa31db3fb4d7d7154feef3fd778318c64a3 Is it possible to have the client/api do the same thing? https://replicate.com/docs/reference/http#predictions.list doesn't seem to have the...