utils.py 26.8 KB
Newer Older
comfyanonymous's avatar
comfyanonymous committed
1
import torch
comfyanonymous's avatar
comfyanonymous committed
2
import math
3
import struct
4
import comfy.checkpoint_pickle
5
import safetensors.torch
comfyanonymous's avatar
comfyanonymous committed
6
import numpy as np
7
from PIL import Image
8
import logging
9
import itertools
comfyanonymous's avatar
comfyanonymous committed
10

comfyanonymous's avatar
comfyanonymous committed
11
12
13
def load_torch_file(ckpt, safe_load=False, device=None):
    if device is None:
        device = torch.device("cpu")
14
    if ckpt.lower().endswith(".safetensors"):
comfyanonymous's avatar
comfyanonymous committed
15
        sd = safetensors.torch.load_file(ckpt, device=device.type)
16
    else:
17
18
        if safe_load:
            if not 'weights_only' in torch.load.__code__.co_varnames:
19
                logging.warning("Warning torch.load doesn't support weights_only on this pytorch version, loading unsafely.")
20
                safe_load = False
21
        if safe_load:
comfyanonymous's avatar
comfyanonymous committed
22
            pl_sd = torch.load(ckpt, map_location=device, weights_only=True)
23
        else:
comfyanonymous's avatar
comfyanonymous committed
24
            pl_sd = torch.load(ckpt, map_location=device, pickle_module=comfy.checkpoint_pickle)
25
        if "global_step" in pl_sd:
comfyanonymous's avatar
comfyanonymous committed
26
            logging.debug(f"Global Step: {pl_sd['global_step']}")
27
28
29
30
31
32
        if "state_dict" in pl_sd:
            sd = pl_sd["state_dict"]
        else:
            sd = pl_sd
    return sd

33
34
35
36
37
38
def save_torch_file(sd, ckpt, metadata=None):
    if metadata is not None:
        safetensors.torch.save_file(sd, ckpt, metadata=metadata)
    else:
        safetensors.torch.save_file(sd, ckpt)

39
40
41
42
43
44
45
def calculate_parameters(sd, prefix=""):
    params = 0
    for k in sd.keys():
        if k.startswith(prefix):
            params += sd[k].nelement()
    return params

46
47
48
49
50
51
def state_dict_key_replace(state_dict, keys_to_replace):
    for x in keys_to_replace:
        if x in state_dict:
            state_dict[keys_to_replace[x]] = state_dict.pop(x)
    return state_dict

comfyanonymous's avatar
comfyanonymous committed
52
53
54
55
56
def state_dict_prefix_replace(state_dict, replace_prefix, filter_keys=False):
    if filter_keys:
        out = {}
    else:
        out = state_dict
57
58
59
    for rp in replace_prefix:
        replace = list(map(lambda a: (a, "{}{}".format(replace_prefix[rp], a[len(rp):])), filter(lambda a: a.startswith(rp), state_dict.keys())))
        for x in replace:
comfyanonymous's avatar
comfyanonymous committed
60
61
62
            w = state_dict.pop(x[0])
            out[x[1]] = w
    return out
63
64


65
def transformers_convert(sd, prefix_from, prefix_to, number):
comfyanonymous's avatar
comfyanonymous committed
66
    keys_to_replace = {
67
68
69
70
        "{}positional_embedding": "{}embeddings.position_embedding.weight",
        "{}token_embedding.weight": "{}embeddings.token_embedding.weight",
        "{}ln_final.weight": "{}final_layer_norm.weight",
        "{}ln_final.bias": "{}final_layer_norm.bias",
comfyanonymous's avatar
comfyanonymous committed
71
72
73
74
75
76
77
    }

    for k in keys_to_replace:
        x = k.format(prefix_from)
        if x in sd:
            sd[keys_to_replace[k].format(prefix_to)] = sd.pop(x)

78
79
80
81
82
83
84
85
86
87
88
    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"]:
89
90
                k = "{}transformer.resblocks.{}.{}.{}".format(prefix_from, resblock, x, y)
                k_to = "{}encoder.layers.{}.{}.{}".format(prefix_to, resblock, resblock_to_replace[x], y)
91
92
93
94
                if k in sd:
                    sd[k_to] = sd.pop(k)

        for y in ["weight", "bias"]:
95
            k_from = "{}transformer.resblocks.{}.attn.in_proj_{}".format(prefix_from, resblock, y)
96
97
98
99
100
            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"]
101
                    k_to = "{}encoder.layers.{}.{}.{}".format(prefix_to, resblock, p[x], y)
102
                    sd[k_to] = weights[shape_from*x:shape_from*(x + 1)]
103
104
105
106
107
108
109
110
111
112
113
114

    return sd

def clip_text_transformers_convert(sd, prefix_from, prefix_to):
    sd = transformers_convert(sd, prefix_from, "{}text_model.".format(prefix_to), 32)

    tp = "{}text_projection.weight".format(prefix_from)
    if tp in sd:
        sd["{}text_projection.weight".format(prefix_to)] = sd.pop(tp)

    tp = "{}text_projection".format(prefix_from)
    if tp in sd:
115
        sd["{}text_projection.weight".format(prefix_to)] = sd.pop(tp).transpose(0, 1).contiguous()
116
117
    return sd

118

119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
UNET_MAP_ATTENTIONS = {
    "proj_in.weight",
    "proj_in.bias",
    "proj_out.weight",
    "proj_out.bias",
    "norm.weight",
    "norm.bias",
}

TRANSFORMER_BLOCKS = {
    "norm1.weight",
    "norm1.bias",
    "norm2.weight",
    "norm2.bias",
    "norm3.weight",
    "norm3.bias",
    "attn1.to_q.weight",
    "attn1.to_k.weight",
    "attn1.to_v.weight",
    "attn1.to_out.0.weight",
    "attn1.to_out.0.bias",
    "attn2.to_q.weight",
    "attn2.to_k.weight",
    "attn2.to_v.weight",
    "attn2.to_out.0.weight",
    "attn2.to_out.0.bias",
    "ff.net.0.proj.weight",
    "ff.net.0.proj.bias",
    "ff.net.2.weight",
    "ff.net.2.bias",
}

UNET_MAP_RESNET = {
    "in_layers.2.weight": "conv1.weight",
    "in_layers.2.bias": "conv1.bias",
    "emb_layers.1.weight": "time_emb_proj.weight",
    "emb_layers.1.bias": "time_emb_proj.bias",
    "out_layers.3.weight": "conv2.weight",
    "out_layers.3.bias": "conv2.bias",
    "skip_connection.weight": "conv_shortcut.weight",
    "skip_connection.bias": "conv_shortcut.bias",
    "in_layers.0.weight": "norm1.weight",
    "in_layers.0.bias": "norm1.bias",
    "out_layers.0.weight": "norm2.weight",
    "out_layers.0.bias": "norm2.bias",
}

166
UNET_MAP_BASIC = {
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
    ("label_emb.0.0.weight", "class_embedding.linear_1.weight"),
    ("label_emb.0.0.bias", "class_embedding.linear_1.bias"),
    ("label_emb.0.2.weight", "class_embedding.linear_2.weight"),
    ("label_emb.0.2.bias", "class_embedding.linear_2.bias"),
    ("label_emb.0.0.weight", "add_embedding.linear_1.weight"),
    ("label_emb.0.0.bias", "add_embedding.linear_1.bias"),
    ("label_emb.0.2.weight", "add_embedding.linear_2.weight"),
    ("label_emb.0.2.bias", "add_embedding.linear_2.bias"),
    ("input_blocks.0.0.weight", "conv_in.weight"),
    ("input_blocks.0.0.bias", "conv_in.bias"),
    ("out.0.weight", "conv_norm_out.weight"),
    ("out.0.bias", "conv_norm_out.bias"),
    ("out.2.weight", "conv_out.weight"),
    ("out.2.bias", "conv_out.bias"),
    ("time_embed.0.weight", "time_embedding.linear_1.weight"),
    ("time_embed.0.bias", "time_embedding.linear_1.bias"),
    ("time_embed.2.weight", "time_embedding.linear_2.weight"),
    ("time_embed.2.bias", "time_embedding.linear_2.bias")
185
186
}

187
def unet_to_diffusers(unet_config):
comfyanonymous's avatar
comfyanonymous committed
188
189
    if "num_res_blocks" not in unet_config:
        return {}
190
191
    num_res_blocks = unet_config["num_res_blocks"]
    channel_mult = unet_config["channel_mult"]
192
193
    transformer_depth = unet_config["transformer_depth"][:]
    transformer_depth_output = unet_config["transformer_depth_output"][:]
194
    num_blocks = len(channel_mult)
195
196

    transformers_mid = unet_config.get("transformer_depth_middle", None)
197
198
199
200
201
202
203

    diffusers_unet_map = {}
    for x in range(num_blocks):
        n = 1 + (num_res_blocks[x] + 1) * x
        for i in range(num_res_blocks[x]):
            for b in UNET_MAP_RESNET:
                diffusers_unet_map["down_blocks.{}.resnets.{}.{}".format(x, i, UNET_MAP_RESNET[b])] = "input_blocks.{}.0.{}".format(n, b)
204
205
            num_transformers = transformer_depth.pop(0)
            if num_transformers > 0:
206
207
                for b in UNET_MAP_ATTENTIONS:
                    diffusers_unet_map["down_blocks.{}.attentions.{}.{}".format(x, i, b)] = "input_blocks.{}.1.{}".format(n, b)
208
                for t in range(num_transformers):
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
                    for b in TRANSFORMER_BLOCKS:
                        diffusers_unet_map["down_blocks.{}.attentions.{}.transformer_blocks.{}.{}".format(x, i, t, b)] = "input_blocks.{}.1.transformer_blocks.{}.{}".format(n, t, b)
            n += 1
        for k in ["weight", "bias"]:
            diffusers_unet_map["down_blocks.{}.downsamplers.0.conv.{}".format(x, k)] = "input_blocks.{}.0.op.{}".format(n, k)

    i = 0
    for b in UNET_MAP_ATTENTIONS:
        diffusers_unet_map["mid_block.attentions.{}.{}".format(i, b)] = "middle_block.1.{}".format(b)
    for t in range(transformers_mid):
        for b in TRANSFORMER_BLOCKS:
            diffusers_unet_map["mid_block.attentions.{}.transformer_blocks.{}.{}".format(i, t, b)] = "middle_block.1.transformer_blocks.{}.{}".format(t, b)

    for i, n in enumerate([0, 2]):
        for b in UNET_MAP_RESNET:
            diffusers_unet_map["mid_block.resnets.{}.{}".format(i, UNET_MAP_RESNET[b])] = "middle_block.{}.{}".format(n, b)

    num_res_blocks = list(reversed(num_res_blocks))
    for x in range(num_blocks):
        n = (num_res_blocks[x] + 1) * x
        l = num_res_blocks[x] + 1
        for i in range(l):
            c = 0
            for b in UNET_MAP_RESNET:
                diffusers_unet_map["up_blocks.{}.resnets.{}.{}".format(x, i, UNET_MAP_RESNET[b])] = "output_blocks.{}.0.{}".format(n, b)
            c += 1
235
236
            num_transformers = transformer_depth_output.pop()
            if num_transformers > 0:
237
238
239
                c += 1
                for b in UNET_MAP_ATTENTIONS:
                    diffusers_unet_map["up_blocks.{}.attentions.{}.{}".format(x, i, b)] = "output_blocks.{}.1.{}".format(n, b)
240
                for t in range(num_transformers):
241
242
243
244
245
246
                    for b in TRANSFORMER_BLOCKS:
                        diffusers_unet_map["up_blocks.{}.attentions.{}.transformer_blocks.{}.{}".format(x, i, t, b)] = "output_blocks.{}.1.transformer_blocks.{}.{}".format(n, t, b)
            if i == l - 1:
                for k in ["weight", "bias"]:
                    diffusers_unet_map["up_blocks.{}.upsamplers.0.conv.{}".format(x, k)] = "output_blocks.{}.{}.conv.{}".format(n, c, k)
            n += 1
247
248

    for k in UNET_MAP_BASIC:
249
        diffusers_unet_map[k[1]] = k[0]
250

251
252
    return diffusers_unet_map

253
254
255
256
257
def swap_scale_shift(weight):
    shift, scale = weight.chunk(2, dim=0)
    new_weight = torch.cat([scale, shift], dim=0)
    return new_weight

comfyanonymous's avatar
comfyanonymous committed
258
259
260
261
262
263
264
265
266
267
268
269
270
271
MMDIT_MAP_BASIC = {
    ("context_embedder.bias", "context_embedder.bias"),
    ("context_embedder.weight", "context_embedder.weight"),
    ("t_embedder.mlp.0.bias", "time_text_embed.timestep_embedder.linear_1.bias"),
    ("t_embedder.mlp.0.weight", "time_text_embed.timestep_embedder.linear_1.weight"),
    ("t_embedder.mlp.2.bias", "time_text_embed.timestep_embedder.linear_2.bias"),
    ("t_embedder.mlp.2.weight", "time_text_embed.timestep_embedder.linear_2.weight"),
    ("x_embedder.proj.bias", "pos_embed.proj.bias"),
    ("x_embedder.proj.weight", "pos_embed.proj.weight"),
    ("y_embedder.mlp.0.bias", "time_text_embed.text_embedder.linear_1.bias"),
    ("y_embedder.mlp.0.weight", "time_text_embed.text_embedder.linear_1.weight"),
    ("y_embedder.mlp.2.bias", "time_text_embed.text_embedder.linear_2.bias"),
    ("y_embedder.mlp.2.weight", "time_text_embed.text_embedder.linear_2.weight"),
    ("pos_embed", "pos_embed.pos_embed"),
272
273
    ("final_layer.adaLN_modulation.1.bias", "norm_out.linear.bias", swap_scale_shift),
    ("final_layer.adaLN_modulation.1.weight", "norm_out.linear.weight", swap_scale_shift),
comfyanonymous's avatar
comfyanonymous committed
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
    ("final_layer.linear.bias", "proj_out.bias"),
    ("final_layer.linear.weight", "proj_out.weight"),
}

MMDIT_MAP_BLOCK = {
    ("context_block.adaLN_modulation.1.bias", "norm1_context.linear.bias"),
    ("context_block.adaLN_modulation.1.weight", "norm1_context.linear.weight"),
    ("context_block.attn.proj.bias", "attn.to_add_out.bias"),
    ("context_block.attn.proj.weight", "attn.to_add_out.weight"),
    ("context_block.mlp.fc1.bias", "ff_context.net.0.proj.bias"),
    ("context_block.mlp.fc1.weight", "ff_context.net.0.proj.weight"),
    ("context_block.mlp.fc2.bias", "ff_context.net.2.bias"),
    ("context_block.mlp.fc2.weight", "ff_context.net.2.weight"),
    ("x_block.adaLN_modulation.1.bias", "norm1.linear.bias"),
    ("x_block.adaLN_modulation.1.weight", "norm1.linear.weight"),
    ("x_block.attn.proj.bias", "attn.to_out.0.bias"),
    ("x_block.attn.proj.weight", "attn.to_out.0.weight"),
    ("x_block.mlp.fc1.bias", "ff.net.0.proj.bias"),
    ("x_block.mlp.fc1.weight", "ff.net.0.proj.weight"),
    ("x_block.mlp.fc2.bias", "ff.net.2.bias"),
    ("x_block.mlp.fc2.weight", "ff.net.2.weight"),
}

def mmdit_to_diffusers(mmdit_config, output_prefix=""):
    key_map = {}

    depth = mmdit_config.get("depth", 0)
301
302
    num_blocks = mmdit_config.get("num_blocks", depth)
    for i in range(num_blocks):
comfyanonymous's avatar
comfyanonymous committed
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
        block_from = "transformer_blocks.{}".format(i)
        block_to = "{}joint_blocks.{}".format(output_prefix, i)

        offset = depth * 64

        for end in ("weight", "bias"):
            k = "{}.attn.".format(block_from)
            qkv = "{}.x_block.attn.qkv.{}".format(block_to, end)
            key_map["{}to_q.{}".format(k, end)] = (qkv, (0, 0, offset))
            key_map["{}to_k.{}".format(k, end)] = (qkv, (0, offset, offset))
            key_map["{}to_v.{}".format(k, end)] = (qkv, (0, offset * 2, offset))

            qkv = "{}.context_block.attn.qkv.{}".format(block_to, end)
            key_map["{}add_q_proj.{}".format(k, end)] = (qkv, (0, 0, offset))
            key_map["{}add_k_proj.{}".format(k, end)] = (qkv, (0, offset, offset))
            key_map["{}add_v_proj.{}".format(k, end)] = (qkv, (0, offset * 2, offset))

        for k in MMDIT_MAP_BLOCK:
            key_map["{}.{}".format(block_from, k[1])] = "{}.{}".format(block_to, k[0])

323
324
325
326
327
328
329
330
331
    map_basic = MMDIT_MAP_BASIC.copy()
    map_basic.add(("joint_blocks.{}.context_block.adaLN_modulation.1.bias".format(depth - 1), "transformer_blocks.{}.norm1_context.linear.bias".format(depth - 1), swap_scale_shift))
    map_basic.add(("joint_blocks.{}.context_block.adaLN_modulation.1.weight".format(depth - 1), "transformer_blocks.{}.norm1_context.linear.weight".format(depth - 1), swap_scale_shift))

    for k in map_basic:
        if len(k) > 2:
            key_map[k[1]] = ("{}{}".format(output_prefix, k[0]), None, k[2])
        else:
            key_map[k[1]] = "{}{}".format(output_prefix, k[0])
comfyanonymous's avatar
comfyanonymous committed
332
333
334

    return key_map

335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404

def auraflow_to_diffusers(mmdit_config, output_prefix=""):
    n_double_layers = mmdit_config.get("n_double_layers", 0)
    n_layers = mmdit_config.get("n_layers", 0)

    key_map = {}
    for i in range(n_layers):
        if i < n_double_layers:
            index = i
            prefix_from = "joint_transformer_blocks"
            prefix_to = "{}double_layers".format(output_prefix)
            block_map = {
                            "attn.to_q.weight": "attn.w2q.weight",
                            "attn.to_k.weight": "attn.w2k.weight",
                            "attn.to_v.weight": "attn.w2v.weight",
                            "attn.to_out.0.weight": "attn.w2o.weight",
                            "attn.add_q_proj.weight": "attn.w1q.weight",
                            "attn.add_k_proj.weight": "attn.w1k.weight",
                            "attn.add_v_proj.weight": "attn.w1v.weight",
                            "attn.to_add_out.weight": "attn.w1o.weight",
                            "ff.linear_1.weight": "mlpX.c_fc1.weight",
                            "ff.linear_2.weight": "mlpX.c_fc2.weight",
                            "ff.out_projection.weight": "mlpX.c_proj.weight",
                            "ff_context.linear_1.weight": "mlpC.c_fc1.weight",
                            "ff_context.linear_2.weight": "mlpC.c_fc2.weight",
                            "ff_context.out_projection.weight": "mlpC.c_proj.weight",
                            "norm1.linear.weight": "modX.1.weight",
                            "norm1_context.linear.weight": "modC.1.weight",
                        }
        else:
            index = i - n_double_layers
            prefix_from = "single_transformer_blocks"
            prefix_to = "{}single_layers".format(output_prefix)

            block_map = {
                            "attn.to_q.weight": "attn.w1q.weight",
                            "attn.to_k.weight": "attn.w1k.weight",
                            "attn.to_v.weight": "attn.w1v.weight",
                            "attn.to_out.0.weight": "attn.w1o.weight",
                            "norm1.linear.weight": "modCX.1.weight",
                            "ff.linear_1.weight": "mlp.c_fc1.weight",
                            "ff.linear_2.weight": "mlp.c_fc2.weight",
                            "ff.out_projection.weight": "mlp.c_proj.weight"
                        }

        for k in block_map:
            key_map["{}.{}.{}".format(prefix_from, index, k)] = "{}.{}.{}".format(prefix_to, index, block_map[k])

    MAP_BASIC = {
        ("positional_encoding", "pos_embed.pos_embed"),
        ("register_tokens", "register_tokens"),
        ("t_embedder.mlp.0.weight", "time_step_proj.linear_1.weight"),
        ("t_embedder.mlp.0.bias", "time_step_proj.linear_1.bias"),
        ("t_embedder.mlp.2.weight", "time_step_proj.linear_2.weight"),
        ("t_embedder.mlp.2.bias", "time_step_proj.linear_2.bias"),
        ("cond_seq_linear.weight", "context_embedder.weight"),
        ("init_x_linear.weight", "pos_embed.proj.weight"),
        ("init_x_linear.bias", "pos_embed.proj.bias"),
        ("final_linear.weight", "proj_out.weight"),
        ("modF.1.weight", "norm_out.linear.weight", swap_scale_shift),
    }

    for k in MAP_BASIC:
        if len(k) > 2:
            key_map[k[1]] = ("{}{}".format(output_prefix, k[0]), None, k[2])
        else:
            key_map[k[1]] = "{}{}".format(output_prefix, k[0])

    return key_map

405
406
407
408
409
def repeat_to_batch_size(tensor, batch_size, dim=0):
    if tensor.shape[dim] > batch_size:
        return tensor.narrow(dim, 0, batch_size)
    elif tensor.shape[dim] < batch_size:
        return tensor.repeat(dim * [1] + [math.ceil(batch_size / tensor.shape[dim])] + [1] * (len(tensor.shape) - 1 - dim)).narrow(dim, 0, batch_size)
410
411
    return tensor

412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
def resize_to_batch_size(tensor, batch_size):
    in_batch_size = tensor.shape[0]
    if in_batch_size == batch_size:
        return tensor

    if batch_size <= 1:
        return tensor[:batch_size]

    output = torch.empty([batch_size] + list(tensor.shape)[1:], dtype=tensor.dtype, device=tensor.device)
    if batch_size < in_batch_size:
        scale = (in_batch_size - 1) / (batch_size - 1)
        for i in range(batch_size):
            output[i] = tensor[min(round(i * scale), in_batch_size - 1)]
    else:
        scale = in_batch_size / batch_size
        for i in range(batch_size):
            output[i] = tensor[min(math.floor((i + 0.5) * scale), in_batch_size - 1)]

    return output

432
433
434
435
436
437
def convert_sd_to(state_dict, dtype):
    keys = list(state_dict.keys())
    for k in keys:
        state_dict[k] = state_dict[k].to(dtype)
    return state_dict

438
439
440
441
442
443
444
445
def safetensors_header(safetensors_path, max_size=100*1024*1024):
    with open(safetensors_path, "rb") as f:
        header = f.read(8)
        length_of_header = struct.unpack('<Q', header)[0]
        if length_of_header > max_size:
            return None
        return f.read(length_of_header)

446
447
448
449
450
def set_attr(obj, attr, value):
    attrs = attr.split(".")
    for name in attrs[:-1]:
        obj = getattr(obj, name)
    prev = getattr(obj, attrs[-1])
451
452
453
454
455
    setattr(obj, attrs[-1], value)
    return prev

def set_attr_param(obj, attr, value):
    return set_attr(obj, attr, torch.nn.Parameter(value, requires_grad=False))
456

457
458
459
460
461
462
463
464
def copy_to_param(obj, attr, value):
    # inplace update tensor instead of replacing it
    attrs = attr.split(".")
    for name in attrs[:-1]:
        obj = getattr(obj, name)
    prev = getattr(obj, attrs[-1])
    prev.data.copy_(value)

465
466
467
468
469
470
def get_attr(obj, attr):
    attrs = attr.split(".")
    for name in attrs:
        obj = getattr(obj, name)
    return obj

471
def bislerp(samples, width, height):
BlenderNeko's avatar
BlenderNeko committed
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
    def slerp(b1, b2, r):
        '''slerps batches b1, b2 according to ratio r, batches should be flat e.g. NxC'''
        
        c = b1.shape[-1]

        #norms
        b1_norms = torch.norm(b1, dim=-1, keepdim=True)
        b2_norms = torch.norm(b2, dim=-1, keepdim=True)

        #normalize
        b1_normalized = b1 / b1_norms
        b2_normalized = b2 / b2_norms

        #zero when norms are zero
        b1_normalized[b1_norms.expand(-1,c) == 0.0] = 0.0
        b2_normalized[b2_norms.expand(-1,c) == 0.0] = 0.0

        #slerp
        dot = (b1_normalized*b2_normalized).sum(1)
        omega = torch.acos(dot)
492
        so = torch.sin(omega)
BlenderNeko's avatar
BlenderNeko committed
493
494
495
496
497
498
499
500
501
502

        #technically not mathematically correct, but more pleasing?
        res = (torch.sin((1.0-r.squeeze(1))*omega)/so).unsqueeze(1)*b1_normalized + (torch.sin(r.squeeze(1)*omega)/so).unsqueeze(1) * b2_normalized
        res *= (b1_norms * (1.0-r) + b2_norms * r).expand(-1,c)

        #edge cases for same or polar opposites
        res[dot > 1 - 1e-5] = b1[dot > 1 - 1e-5] 
        res[dot < 1e-5 - 1] = (b1 * (1.0-r) + b2 * r)[dot < 1e-5 - 1]
        return res
    
comfyanonymous's avatar
comfyanonymous committed
503
504
    def generate_bilinear_data(length_old, length_new, device):
        coords_1 = torch.arange(length_old, dtype=torch.float32, device=device).reshape((1,1,1,-1))
BlenderNeko's avatar
BlenderNeko committed
505
506
507
508
        coords_1 = torch.nn.functional.interpolate(coords_1, size=(1, length_new), mode="bilinear")
        ratios = coords_1 - coords_1.floor()
        coords_1 = coords_1.to(torch.int64)
        
comfyanonymous's avatar
comfyanonymous committed
509
        coords_2 = torch.arange(length_old, dtype=torch.float32, device=device).reshape((1,1,1,-1)) + 1
BlenderNeko's avatar
BlenderNeko committed
510
511
512
513
        coords_2[:,:,:,-1] -= 1
        coords_2 = torch.nn.functional.interpolate(coords_2, size=(1, length_new), mode="bilinear")
        coords_2 = coords_2.to(torch.int64)
        return ratios, coords_1, coords_2
514
515
516

    orig_dtype = samples.dtype
    samples = samples.float()
BlenderNeko's avatar
BlenderNeko committed
517
518
519
    n,c,h,w = samples.shape
    h_new, w_new = (height, width)
    
520
    #linear w
comfyanonymous's avatar
comfyanonymous committed
521
    ratios, coords_1, coords_2 = generate_bilinear_data(w, w_new, samples.device)
522
523
524
    coords_1 = coords_1.expand((n, c, h, -1))
    coords_2 = coords_2.expand((n, c, h, -1))
    ratios = ratios.expand((n, 1, h, -1))
BlenderNeko's avatar
BlenderNeko committed
525

comfyanonymous's avatar
comfyanonymous committed
526
527
528
    pass_1 = samples.gather(-1,coords_1).movedim(1, -1).reshape((-1,c))
    pass_2 = samples.gather(-1,coords_2).movedim(1, -1).reshape((-1,c))
    ratios = ratios.movedim(1, -1).reshape((-1,1))
BlenderNeko's avatar
BlenderNeko committed
529
530

    result = slerp(pass_1, pass_2, ratios)
comfyanonymous's avatar
comfyanonymous committed
531
    result = result.reshape(n, h, w_new, c).movedim(-1, 1)
BlenderNeko's avatar
BlenderNeko committed
532

533
    #linear h
comfyanonymous's avatar
comfyanonymous committed
534
    ratios, coords_1, coords_2 = generate_bilinear_data(h, h_new, samples.device)
535
536
537
    coords_1 = coords_1.reshape((1,1,-1,1)).expand((n, c, -1, w_new))
    coords_2 = coords_2.reshape((1,1,-1,1)).expand((n, c, -1, w_new))
    ratios = ratios.reshape((1,1,-1,1)).expand((n, 1, -1, w_new))
BlenderNeko's avatar
BlenderNeko committed
538

comfyanonymous's avatar
comfyanonymous committed
539
540
541
    pass_1 = result.gather(-2,coords_1).movedim(1, -1).reshape((-1,c))
    pass_2 = result.gather(-2,coords_2).movedim(1, -1).reshape((-1,c))
    ratios = ratios.movedim(1, -1).reshape((-1,1))
BlenderNeko's avatar
BlenderNeko committed
542
543

    result = slerp(pass_1, pass_2, ratios)
comfyanonymous's avatar
comfyanonymous committed
544
    result = result.reshape(n, h_new, w_new, c).movedim(-1, 1)
545
    return result.to(orig_dtype)
546

547
def lanczos(samples, width, height):
comfyanonymous's avatar
comfyanonymous committed
548
    images = [Image.fromarray(np.clip(255. * image.movedim(0, -1).cpu().numpy(), 0, 255).astype(np.uint8)) for image in samples]
549
    images = [image.resize((width, height), resample=Image.Resampling.LANCZOS) for image in images]
comfyanonymous's avatar
comfyanonymous committed
550
    images = [torch.from_numpy(np.array(image).astype(np.float32) / 255.0).movedim(-1, 0) for image in images]
551
    result = torch.stack(images)
552
    return result.to(samples.device, samples.dtype)
553

comfyanonymous's avatar
comfyanonymous committed
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
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
569
570
571

        if upscale_method == "bislerp":
            return bislerp(s, width, height)
572
573
        elif upscale_method == "lanczos":
            return lanczos(s, width, height)
574
575
        else:
            return torch.nn.functional.interpolate(s, size=(height, width), mode=upscale_method)
576

pythongosssss's avatar
pythongosssss committed
577
def get_tiled_scale_steps(width, height, tile_x, tile_y, overlap):
comfyanonymous's avatar
comfyanonymous committed
578
    return math.ceil((height / (tile_y - overlap))) * math.ceil((width / (tile_x - overlap)))
pythongosssss's avatar
pythongosssss committed
579

580
@torch.inference_mode()
581
582
583
584
def tiled_scale_multidim(samples, function, tile=(64, 64), overlap = 8, upscale_amount = 4, out_channels = 3, output_device="cpu", pbar = None):
    dims = len(tile)
    output = torch.empty([samples.shape[0], out_channels] + list(map(lambda a: round(a * upscale_amount), samples.shape[2:])), device=output_device)

585
586
    for b in range(samples.shape[0]):
        s = samples[b:b+1]
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
        out = torch.zeros([s.shape[0], out_channels] + list(map(lambda a: round(a * upscale_amount), s.shape[2:])), device=output_device)
        out_div = torch.zeros([s.shape[0], out_channels] + list(map(lambda a: round(a * upscale_amount), s.shape[2:])), device=output_device)

        for it in itertools.product(*map(lambda a: range(0, a[0], a[1] - overlap), zip(s.shape[2:], tile))):
            s_in = s
            upscaled = []

            for d in range(dims):
                pos = max(0, min(s.shape[d + 2] - overlap, it[d]))
                l = min(tile[d], s.shape[d + 2] - pos)
                s_in = s_in.narrow(d + 2, pos, l)
                upscaled.append(round(pos * upscale_amount))
            ps = function(s_in).to(output_device)
            mask = torch.ones_like(ps)
            feather = round(overlap * upscale_amount)
            for t in range(feather):
                for d in range(2, dims + 2):
                    m = mask.narrow(d, t, 1)
                    m *= ((1.0/feather) * (t + 1))
                    m = mask.narrow(d, mask.shape[d] -1 -t, 1)
                    m *= ((1.0/feather) * (t + 1))

            o = out
            o_d = out_div
            for d in range(dims):
                o = o.narrow(d + 2, upscaled[d], mask.shape[d + 2])
                o_d = o_d.narrow(d + 2, upscaled[d], mask.shape[d + 2])

            o += ps * mask
            o_d += mask

            if pbar is not None:
                pbar.update(1)
620
621
622

        output[b:b+1] = out/out_div
    return output
623

624
625
626
def tiled_scale(samples, function, tile_x=64, tile_y=64, overlap = 8, upscale_amount = 4, out_channels = 3, output_device="cpu", pbar = None):
    return tiled_scale_multidim(samples, function, (tile_y, tile_x), overlap, upscale_amount, out_channels, output_device, pbar)

627
628
629
630
PROGRESS_BAR_ENABLED = True
def set_progress_bar_enabled(enabled):
    global PROGRESS_BAR_ENABLED
    PROGRESS_BAR_ENABLED = enabled
631
632
633
634
635
636
637
638
639
640
641
642
643

PROGRESS_BAR_HOOK = None
def set_progress_bar_global_hook(function):
    global PROGRESS_BAR_HOOK
    PROGRESS_BAR_HOOK = function

class ProgressBar:
    def __init__(self, total):
        global PROGRESS_BAR_HOOK
        self.total = total
        self.current = 0
        self.hook = PROGRESS_BAR_HOOK

space-nuko's avatar
space-nuko committed
644
    def update_absolute(self, value, total=None, preview=None):
645
646
        if total is not None:
            self.total = total
647
648
649
650
        if value > self.total:
            value = self.total
        self.current = value
        if self.hook is not None:
space-nuko's avatar
space-nuko committed
651
            self.hook(self.current, self.total, preview)
652
653
654

    def update(self, value):
        self.update_absolute(self.current + value)