Could you tell us the same commands you used in the replicate.com model?

[wes-kay]

Trying to get similar results but can't tune them enough to get them.

python sample.py --init_image imgName --skip_timesteps 10 --model_path finetune.pt --batch_size 8 --num_batches 8 --text text

[afiaka87]

@wes-kay The Replicate models use either erlich.pt or ongo.pt

[afiaka87]

@wes-kay I can also add instructions for how to build your own cog image (backend for replicate) so you can have the exact same setup. Would that be useful?

[wes-kay]

@afiaka87 That would be amazing, I was trying to copy the step size of 100 as one of the arguments but I'm seeing that there's no similarities between what's set in the sample.py and the site fields, once I removed that I'm getting pretty similar results as to the site.

I think being able to see what has been set on would be great, as there's not a lot of info I've gotten from reading the code that transfers to the site fields.

parser.add_argument('--model_path', type=str, default='finetune.pt',
                    help='path to the diffusion model')

parser.add_argument('--kl_path', type=str, default='kl-f8.pt',
                    help='path to the LDM first stage model')

parser.add_argument('--bert_path', type=str, default='bert.pt',
                    help='path to the LDM first stage model')

parser.add_argument('--text', type=str, required=False, default='',
                    help='your text prompt')

parser.add_argument('--edit', type=str, required=False,
                    help='path to the image you want to edit (either an image file or .npy containing a numpy array of the image embeddings)')

parser.add_argument('--edit_x', type=int, required=False, default=0,
                    help='x position of the edit image in the generation frame (need to be multiple of 8)')

parser.add_argument('--edit_y', type=int, required=False, default=0,
                    help='y position of the edit image in the generation frame (need to be multiple of 8)')

parser.add_argument('--edit_width', type=int, required=False, default=0,
                    help='width of the edit image in the generation frame (need to be multiple of 8)')

parser.add_argument('--edit_height', type=int, required=False, default=0,
                    help='height of the edit image in the generation frame (need to be multiple of 8)')

parser.add_argument('--mask', type=str, required=False,
                    help='path to a mask image. white pixels = keep, black pixels = discard. width = image width/8, height = image height/8')

parser.add_argument('--negative', type=str, required=False, default='',
                    help='negative text prompt')

parser.add_argument('--init_image', type=str, required=False, default=None,
                    help='init image to use')

parser.add_argument('--skip_timesteps', type=int, required=False, default=0,
                    help='how many diffusion steps are gonna be skipped')

parser.add_argument('--prefix', type=str, required=False, default='',
                    help='prefix for output files')

parser.add_argument('--num_batches', type=int, default=1, required=False,
                    help='number of batches')

parser.add_argument('--batch_size', type=int, default=1, required=False,
                    help='batch size')

parser.add_argument('--width', type=int, default=256, required=False,
                    help='image size of output (multiple of 8)')

parser.add_argument('--height', type=int, default=256, required=False,
                    help='image size of output (multiple of 8)')

parser.add_argument('--seed', type=int, default=-1, required=False,
                    help='random seed')

parser.add_argument('--guidance_scale', type=float, default=5.0, required=False,
                    help='classifier-free guidance scale')

parser.add_argument('--steps', type=int, default=0, required=False,
                    help='number of diffusion steps')

parser.add_argument('--cpu', dest='cpu', action='store_true')

parser.add_argument('--clip_score', dest='clip_score', action='store_true')

parser.add_argument('--clip_guidance',
                    dest='clip_guidance', action='store_true')

parser.add_argument('--clip_guidance_scale', type=float, default=150, required=False,
                    help='Controls how much the image should look like the prompt')  # may need to use lower value for ddim

parser.add_argument('--cutn', type=int, default=16, required=False,
                    help='Number of cuts')

#turn on to use 50 step ddim
parser.add_argument('--ddim', dest='ddim', action='store_true')

#turn on to use 50 step ddim
parser.add_argument('--ddpm', dest='ddpm', action='store_true')

[neurofuzzy]

I would love to have this as well. The replicate setup is perfect for my tastes. I tried putting together a Google Colab but I've never done it before and am a bit lost. There's a colab that uses ongo called "Latent Majesty Diffusion" which gets pretty close but I haven't gotten anything near as good as what comes out of replicate.

[afiaka87]

@wes-kay @neurofuzzy I've added instructions for this btw