bmaltais / dehydrate Goto Github PK

Not sure what the original script did since it had a default alpha value of 0.

parser.add_argument("--str", type=float, help="Strength of the rehydration (-0.05..0.05)", default=0, required=False)`
alpha = args.str
...
for key in tqdm(theta_0.keys(), desc="Stage 1/2?: merge common keys"):
    if "model" in key and key in theta_1:
        theta_0[key] = theta_0[key] * (1 + alpha) + theta_1[key] * (1 - alpha)

Which appears just replaced the final weights with weights that were the same, since theta_0[key] = theta_0[key] * 1 + theta_1[key] * 0 = theta_0[key] = theta_0[key].

Constructing a patch with M1 - M2 = patch and then M1 + patch = M2 does not work due to lossy FP and seems to give me corrupted models.

What I did instead was use thresholding based on the min/max delta of the individual tensors and then apply the most changed values as a patch.

Here's my code: https://gist.github.com/AmericanPresidentJimmyCarter/1947162f371e601ce183070443f41dc2

This results in an about 60 MB patch, but the quality of transfer is not great. You can increase the THRESHOLD_STRENGTH variable but the file sizes increase dramatically while the quality does not seem to. Here are results at THRESHOLD_STRENGTH == 2..

Test code is:

_, extra_data = engine.sample(
    'elon musk in modern disney style',
    4,
    'heun',
    679566949,
    25,
    scale=7.,
)

Using dreambooth'd weights here: https://huggingface.co/nitrosocke/mo-di-diffusion