ComfyUI/comfy/utils.py

import os
import torch
import folder_paths
import numpy as np
from safetensors.torch import save_file, safe_open


def load_torch_file(ckpt, safe_load=False):
    if ckpt.lower().endswith(".safetensors"):
        import safetensors.torch
        sd = safetensors.torch.load_file(ckpt, device="cpu")
    else:
        if safe_load:
            pl_sd = torch.load(ckpt, map_location="cpu", weights_only=True)
        else:
            pl_sd = torch.load(ckpt, map_location="cpu")
        if "global_step" in pl_sd:
            print(f"Global Step: {pl_sd['global_step']}")
        if "state_dict" in pl_sd:
            sd = pl_sd["state_dict"]
        else:
            sd = pl_sd
    return sd


def save_latent(samples, filename):
    filename = os.path.join(folder_paths.get_output_directory(), filename)
    np.save(filename, samples)


def save_attention(attention, filename):
    filename = os.path.join(folder_paths.get_output_directory(), filename)
    save_file({"attention": attention}, filename)
    print(f"Attention tensor saved to {filename}")


def load_attention(filename):
    tensors = {}
    filename = os.path.join(folder_paths.get_output_directory(), filename)
    with safe_open(filename, framework='pt', device=0) as f:
        for k in f.keys():
            tensors[k] = f.get_tensor(k)
    return tensors['attention']


def load_latent(filename):
    filename = os.path.join(folder_paths.get_output_directory(), filename)
    return torch.from_numpy(np.load(filename))

def transformers_convert(sd, prefix_from, prefix_to, number):
    resblock_to_replace = {
        "ln_1": "layer_norm1",
        "ln_2": "layer_norm2",
        "mlp.c_fc": "mlp.fc1",
        "mlp.c_proj": "mlp.fc2",
        "attn.out_proj": "self_attn.out_proj",
    }

    for resblock in range(number):
        for x in resblock_to_replace:
            for y in ["weight", "bias"]:
                k = "{}.transformer.resblocks.{}.{}.{}".format(prefix_from, resblock, x, y)
                k_to = "{}.encoder.layers.{}.{}.{}".format(prefix_to, resblock, resblock_to_replace[x], y)
                if k in sd:
                    sd[k_to] = sd.pop(k)

        for y in ["weight", "bias"]:
            k_from = "{}.transformer.resblocks.{}.attn.in_proj_{}".format(prefix_from, resblock, y)
            if k_from in sd:
                weights = sd.pop(k_from)
                shape_from = weights.shape[0] // 3
                for x in range(3):
                    p = ["self_attn.q_proj", "self_attn.k_proj", "self_attn.v_proj"]
                    k_to = "{}.encoder.layers.{}.{}.{}".format(prefix_to, resblock, p[x], y)
                    sd[k_to] = weights[shape_from*x:shape_from*(x + 1)]
    return sd

def common_upscale(samples, width, height, upscale_method, crop):
        if crop == "center":
            old_width = samples.shape[3]
            old_height = samples.shape[2]
            old_aspect = old_width / old_height
            new_aspect = width / height
            x = 0
            y = 0
            if old_aspect > new_aspect:
                x = round((old_width - old_width * (new_aspect / old_aspect)) / 2)
            elif old_aspect < new_aspect:
                y = round((old_height - old_height * (old_aspect / new_aspect)) / 2)
            s = samples[:,:,y:old_height-y,x:old_width-x]
        else:
            s = samples
        return torch.nn.functional.interpolate(s, size=(height, width), mode=upscale_method)

@torch.inference_mode()
def tiled_scale(samples, function, tile_x=64, tile_y=64, overlap = 8, upscale_amount = 4, out_channels = 3):
    output = torch.empty((samples.shape[0], out_channels, round(samples.shape[2] * upscale_amount), round(samples.shape[3] * upscale_amount)), device="cpu")
    for b in range(samples.shape[0]):
        s = samples[b:b+1]
        out = torch.zeros((s.shape[0], out_channels, round(s.shape[2] * upscale_amount), round(s.shape[3] * upscale_amount)), device="cpu")
        out_div = torch.zeros((s.shape[0], out_channels, round(s.shape[2] * upscale_amount), round(s.shape[3] * upscale_amount)), device="cpu")
        for y in range(0, s.shape[2], tile_y - overlap):
            for x in range(0, s.shape[3], tile_x - overlap):
                s_in = s[:,:,y:y+tile_y,x:x+tile_x]

                ps = function(s_in).cpu()
                mask = torch.ones_like(ps)
                feather = round(overlap * upscale_amount)
                for t in range(feather):
                        mask[:,:,t:1+t,:] *= ((1.0/feather) * (t + 1))
                        mask[:,:,mask.shape[2] -1 -t: mask.shape[2]-t,:] *= ((1.0/feather) * (t + 1))
                        mask[:,:,:,t:1+t] *= ((1.0/feather) * (t + 1))
                        mask[:,:,:,mask.shape[3]- 1 - t: mask.shape[3]- t] *= ((1.0/feather) * (t + 1))
                out[:,:,round(y*upscale_amount):round((y+tile_y)*upscale_amount),round(x*upscale_amount):round((x+tile_x)*upscale_amount)] += ps * mask
                out_div[:,:,round(y*upscale_amount):round((y+tile_y)*upscale_amount),round(x*upscale_amount):round((x+tile_x)*upscale_amount)] += mask

        output[b:b+1] = out/out_div
    return output