Closed SlZeroth closed 1 year ago
the most important parameter is the alpha. For instance, the following set of images (from celeb-a test set) are all generated with alpha 0.6, all the other hyper-parameter are the same (20 step text only + 30 step subject-conditioned).
Maybe tune that hyperparameter a bit more. Additionally, due to the small scale of FFHQ, some prompts are probably just impossible as they are too far away from the distribution.
You can find the whole test set prediction at url
Try changing the model in fastcomposer/fastcomposer/utils.py to something other than the default runwayml/stable-diffusion-1-5 icbinp_v6 non-cherry picked default setting results with source images that were 172x172 Japanese woodblock print prompting did not cooperate with the aforementioned model.
A man <A> and a man <A> in the snow https://ibb.co/0V766qZ https://ibb.co/Gnds0xW https://ibb.co/3rqHLkm https://ibb.co/Cbh52NP https://ibb.co/B33kfqG https://ibb.co/bHTGdDg https://ibb.co/vDh7wtH https://ibb.co/9cBTg1Q
A man <A> and a woman <A> in the snow https://ibb.co/pnZPLXp https://ibb.co/GRNQd0D https://ibb.co/MhgbRtm https://ibb.co/m4bR87M https://ibb.co/dDz6CbD https://ibb.co/qWXdqqt https://ibb.co/kJHNdTP https://ibb.co/G2Hb2ZF
A man <A> and a woman <A> water color painting https://ibb.co/z6g9yHh https://ibb.co/FWDYjTn https://ibb.co/4mgFNcx https://ibb.co/sPq8SQM https://ibb.co/80cbf9z https://ibb.co/19DYvLD https://ibb.co/mDXt3NG https://ibb.co/Hzm6LRf
A man <A> and a woman <A> standing in the garden https://ibb.co/t8jDFnQ https://ibb.co/bzBLTVt https://ibb.co/svkGT8V https://ibb.co/FxrTwMY https://ibb.co/jTKvQq3 https://ibb.co/rMzyzjq https://ibb.co/KmPgm4T https://ibb.co/p3cp7MY
A woman <A> and a woman <A> gardening, backyard https://ibb.co/GH53wX5 https://ibb.co/5rY5wgk https://ibb.co/9YFyRgS https://ibb.co/yNNkqjP https://ibb.co/Dwy5smy https://ibb.co/xDpD1k4 https://ibb.co/rQh6ZGN https://ibb.co/tc0Fwtj
A woman <A> and a woman <A> in the jungle https://ibb.co/741hMCF https://ibb.co/vhC9wLp https://ibb.co/8rGfWgf https://ibb.co/R9YWGSW https://ibb.co/DQNF4KH https://ibb.co/b1PVjVs https://ibb.co/TMZWG8t https://ibb.co/0DpCgPW
Japanese woodblock print prompting did not cooperate with the aforementioned model.
It probably needs smaller alpha. Some times this results in dissimilar figures but it works reasonably robust (you get at least 1-2 good example among 4).
The following is with default settings (and 0.6 alpha)
full list of predictions on test could be found at URL
Could you explain the method to change the model in more detail? I tried using arguments that seemed appropriate, but it didn't work.
--pretrained_model_name_or_path "D:\Programs_D\_MachineLearning\StableDiffusionModels\Online\CheckPoint\anything-v4.0.ckpt"
OSError: It looks like the config file at 'D:\Programs_D\_MachineLearning\StableDiffusionModels\Online\CheckPoint\anything-v4.0.ckpt' is not a valid JSON file.
The script expects diffusers model, https://huggingface.co/spaces/diffusers/sd-to-diffusers
The sd-to-diffusers appears to be broken atm so it is time to do some hoopjumping in the stable-diffusion-webui environment Using Command Prompt, Activate the env that contains stable-diffusion-webui X:\stable-diffusion-webui\venv\Scripts\activate.bat pip list make note of any installed diffusers version, although when stable-diffusion-webui is started again, diffusers should be reinstalled if required.
pip install -U git+https://github.com/huggingface/diffusers
Download - https://huggingface.co/runwayml/stable-diffusion-v1-5/resolve/main/v1-inference.yaml - Right click and save as Place v1-inference.yaml into X:\stable-diffusion-webui\venv\Lib\site-packages\diffusers Download https://github.com/huggingface/diffusers/archive/refs/heads/main.zip Find the scripts folder within the zip, now drag and drop all the files into X:\stable-diffusion-webui\venv\Lib\site-packages\diffusers
Go back to env activated Command Prompt cd X:\stable-diffusion-webui\venv\Lib\site-packages\diffusers
python convert_original_stable_diffusion_to_diffusers.py --checkpoint_path=X:/stable-diffusion-webui/models/Stable-diffusion/disneyPixarCartoon_v10.safetensors --safetensors --scheduler_type=ddim --dump_path=U:/test/anything-4.5 --original_config_file=X:/stable-diffusion-webui/venv/Lib/site-packages/diffusers/v1-inference.yaml
When finished with conversion(s), if required, reinstall the original diffusers version
pip install diffusers==0.16.1
I used https://github.com/mit-han-lab/fastcomposer to generate a two shot image by inserting two images, but it is difficult to get a successful picture. Is there anything I am mistaken or need to change?