model_detection.py 15.3 KB
Newer Older
1
2
import comfy.supported_models
import comfy.supported_models_base
3
4
5
6
7
8
9
10
11
12
13
14
15
16

def count_blocks(state_dict_keys, prefix_string):
    count = 0
    while True:
        c = False
        for k in state_dict_keys:
            if k.startswith(prefix_string.format(count)):
                c = True
                break
        if c == False:
            break
        count += 1
    return count

17
18
19
20
21
22
23
24
25
26
def calculate_transformer_depth(prefix, state_dict_keys, state_dict):
    context_dim = None
    use_linear_in_transformer = False

    transformer_prefix = prefix + "1.transformer_blocks."
    transformer_keys = sorted(list(filter(lambda a: a.startswith(transformer_prefix), state_dict_keys)))
    if len(transformer_keys) > 0:
        last_transformer_depth = count_blocks(state_dict_keys, transformer_prefix + '{}')
        context_dim = state_dict['{}0.attn2.to_k.weight'.format(transformer_prefix)].shape[1]
        use_linear_in_transformer = len(state_dict['{}1.proj_in.weight'.format(prefix)].shape) == 2
comfyanonymous's avatar
comfyanonymous committed
27
28
        time_stack = '{}1.time_stack.0.attn1.to_q.weight'.format(prefix) in state_dict or '{}1.time_mix_blocks.0.attn1.to_q.weight'.format(prefix) in state_dict
        return last_transformer_depth, context_dim, use_linear_in_transformer, time_stack
29
30
    return None

31
def detect_unet_config(state_dict, key_prefix, dtype):
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
    state_dict_keys = list(state_dict.keys())

    unet_config = {
        "use_checkpoint": False,
        "image_size": 32,
        "out_channels": 4,
        "use_spatial_transformer": True,
        "legacy": False
    }

    y_input = '{}label_emb.0.0.weight'.format(key_prefix)
    if y_input in state_dict_keys:
        unet_config["num_classes"] = "sequential"
        unet_config["adm_in_channels"] = state_dict[y_input].shape[1]
    else:
        unet_config["adm_in_channels"] = None

49
    unet_config["dtype"] = dtype
50
51
52
53
54
55
56
    model_channels = state_dict['{}input_blocks.0.0.weight'.format(key_prefix)].shape[0]
    in_channels = state_dict['{}input_blocks.0.0.weight'.format(key_prefix)].shape[1]

    num_res_blocks = []
    channel_mult = []
    attention_resolutions = []
    transformer_depth = []
57
    transformer_depth_output = []
58
59
60
    context_dim = None
    use_linear_in_transformer = False

comfyanonymous's avatar
comfyanonymous committed
61
    video_model = False
62
63
64
65
66
67
68

    current_res = 1
    count = 0

    last_res_blocks = 0
    last_channel_mult = 0

69
70
    input_block_count = count_blocks(state_dict_keys, '{}input_blocks'.format(key_prefix) + '.{}.')
    for count in range(input_block_count):
71
        prefix = '{}input_blocks.{}.'.format(key_prefix, count)
72
73
        prefix_output = '{}output_blocks.{}.'.format(key_prefix, input_block_count - count - 1)

74
75
76
77
        block_keys = sorted(list(filter(lambda a: a.startswith(prefix), state_dict_keys)))
        if len(block_keys) == 0:
            break

78
79
        block_keys_output = sorted(list(filter(lambda a: a.startswith(prefix_output), state_dict_keys)))

80
81
82
83
84
85
86
        if "{}0.op.weight".format(prefix) in block_keys: #new layer
            num_res_blocks.append(last_res_blocks)
            channel_mult.append(last_channel_mult)

            current_res *= 2
            last_res_blocks = 0
            last_channel_mult = 0
87
88
89
90
91
            out = calculate_transformer_depth(prefix_output, state_dict_keys, state_dict)
            if out is not None:
                transformer_depth_output.append(out[0])
            else:
                transformer_depth_output.append(0)
92
93
94
95
96
97
        else:
            res_block_prefix = "{}0.in_layers.0.weight".format(prefix)
            if res_block_prefix in block_keys:
                last_res_blocks += 1
                last_channel_mult = state_dict["{}0.out_layers.3.weight".format(prefix)].shape[0] // model_channels

98
99
100
101
102
103
                out = calculate_transformer_depth(prefix, state_dict_keys, state_dict)
                if out is not None:
                    transformer_depth.append(out[0])
                    if context_dim is None:
                        context_dim = out[1]
                        use_linear_in_transformer = out[2]
comfyanonymous's avatar
comfyanonymous committed
104
                        video_model = out[3]
105
106
107
108
109
110
111
112
113
114
                else:
                    transformer_depth.append(0)

            res_block_prefix = "{}0.in_layers.0.weight".format(prefix_output)
            if res_block_prefix in block_keys_output:
                out = calculate_transformer_depth(prefix_output, state_dict_keys, state_dict)
                if out is not None:
                    transformer_depth_output.append(out[0])
                else:
                    transformer_depth_output.append(0)
115
116
117
118


    num_res_blocks.append(last_res_blocks)
    channel_mult.append(last_channel_mult)
119
120
121
122
    if "{}middle_block.1.proj_in.weight".format(key_prefix) in state_dict_keys:
        transformer_depth_middle = count_blocks(state_dict_keys, '{}middle_block.1.transformer_blocks.'.format(key_prefix) + '{}')
    else:
        transformer_depth_middle = -1
123
124
125
126
127

    unet_config["in_channels"] = in_channels
    unet_config["model_channels"] = model_channels
    unet_config["num_res_blocks"] = num_res_blocks
    unet_config["transformer_depth"] = transformer_depth
128
    unet_config["transformer_depth_output"] = transformer_depth_output
129
130
131
132
    unet_config["channel_mult"] = channel_mult
    unet_config["transformer_depth_middle"] = transformer_depth_middle
    unet_config['use_linear_in_transformer'] = use_linear_in_transformer
    unet_config["context_dim"] = context_dim
comfyanonymous's avatar
comfyanonymous committed
133
134
135
136
137
138
139
140
141
142
143
144
145

    if video_model:
        unet_config["extra_ff_mix_layer"] = True
        unet_config["use_spatial_context"] = True
        unet_config["merge_strategy"] = "learned_with_images"
        unet_config["merge_factor"] = 0.0
        unet_config["video_kernel_size"] = [3, 1, 1]
        unet_config["use_temporal_resblock"] = True
        unet_config["use_temporal_attention"] = True
    else:
        unet_config["use_temporal_resblock"] = False
        unet_config["use_temporal_attention"] = False

146
147
    return unet_config

148
def model_config_from_unet_config(unet_config):
149
    for model_config in comfy.supported_models.models:
150
151
152
        if model_config.matches(unet_config):
            return model_config(unet_config)

153
    print("no match", unet_config)
154
    return None
155

156
157
def model_config_from_unet(state_dict, unet_key_prefix, dtype, use_base_if_no_match=False):
    unet_config = detect_unet_config(state_dict, unet_key_prefix, dtype)
158
159
160
161
162
    model_config = model_config_from_unet_config(unet_config)
    if model_config is None and use_base_if_no_match:
        return comfy.supported_models_base.BASE(unet_config)
    else:
        return model_config
163

164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
def convert_config(unet_config):
    new_config = unet_config.copy()
    num_res_blocks = new_config.get("num_res_blocks", None)
    channel_mult = new_config.get("channel_mult", None)

    if isinstance(num_res_blocks, int):
        num_res_blocks = len(channel_mult) * [num_res_blocks]

    if "attention_resolutions" in new_config:
        attention_resolutions = new_config.pop("attention_resolutions")
        transformer_depth = new_config.get("transformer_depth", None)
        transformer_depth_middle = new_config.get("transformer_depth_middle", None)

        if isinstance(transformer_depth, int):
            transformer_depth = len(channel_mult) * [transformer_depth]
        if transformer_depth_middle is None:
            transformer_depth_middle =  transformer_depth[-1]
        t_in = []
        t_out = []
        s = 1
        for i in range(len(num_res_blocks)):
            res = num_res_blocks[i]
            d = 0
            if s in attention_resolutions:
                d = transformer_depth[i]

            t_in += [d] * res
            t_out += [d] * (res + 1)
            s *= 2
        transformer_depth = t_in
        transformer_depth_output = t_out
        new_config["transformer_depth"] = t_in
        new_config["transformer_depth_output"] = t_out
        new_config["transformer_depth_middle"] = transformer_depth_middle

    new_config["num_res_blocks"] = num_res_blocks
    return new_config


203
def unet_config_from_diffusers_unet(state_dict, dtype):
204
    match = {}
205
    transformer_depth = []
206
207

    attn_res = 1
208
209
210
211
212
213
214
215
216
    down_blocks = count_blocks(state_dict, "down_blocks.{}")
    for i in range(down_blocks):
        attn_blocks = count_blocks(state_dict, "down_blocks.{}.attentions.".format(i) + '{}')
        for ab in range(attn_blocks):
            transformer_count = count_blocks(state_dict, "down_blocks.{}.attentions.{}.transformer_blocks.".format(i, ab) + '{}')
            transformer_depth.append(transformer_count)
            if transformer_count > 0:
                match["context_dim"] = state_dict["down_blocks.{}.attentions.{}.transformer_blocks.0.attn2.to_k.weight".format(i, ab)].shape[1]

217
        attn_res *= 2
218
219
220
        if attn_blocks == 0:
            transformer_depth.append(0)
            transformer_depth.append(0)
221

222
    match["transformer_depth"] = transformer_depth
223

224
225
226
227
228
229
230
231
232
    match["model_channels"] = state_dict["conv_in.weight"].shape[0]
    match["in_channels"] = state_dict["conv_in.weight"].shape[1]
    match["adm_in_channels"] = None
    if "class_embedding.linear_1.weight" in state_dict:
        match["adm_in_channels"] = state_dict["class_embedding.linear_1.weight"].shape[1]
    elif "add_embedding.linear_1.weight" in state_dict:
        match["adm_in_channels"] = state_dict["add_embedding.linear_1.weight"].shape[1]

    SDXL = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False,
233
            'num_classes': 'sequential', 'adm_in_channels': 2816, 'dtype': dtype, 'in_channels': 4, 'model_channels': 320,
234
235
            'num_res_blocks': [2, 2, 2], 'transformer_depth': [0, 0, 2, 2, 10, 10], 'channel_mult': [1, 2, 4], 'transformer_depth_middle': 10,
            'use_linear_in_transformer': True, 'context_dim': 2048, 'num_head_channels': 64, 'transformer_depth_output': [0, 0, 0, 2, 2, 2, 10, 10, 10]}
236
237

    SDXL_refiner = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False,
238
                    'num_classes': 'sequential', 'adm_in_channels': 2560, 'dtype': dtype, 'in_channels': 4, 'model_channels': 384,
239
240
                    'num_res_blocks': [2, 2, 2, 2], 'transformer_depth': [0, 0, 4, 4, 4, 4, 0, 0], 'channel_mult': [1, 2, 4, 4], 'transformer_depth_middle': 4,
                    'use_linear_in_transformer': True, 'context_dim': 1280, 'num_head_channels': 64, 'transformer_depth_output': [0, 0, 0, 4, 4, 4, 4, 4, 4, 0, 0, 0]}
241
242

    SD21 = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False,
243
244
245
            'adm_in_channels': None, 'dtype': dtype, 'in_channels': 4, 'model_channels': 320, 'num_res_blocks': [2, 2, 2, 2],
            'transformer_depth': [1, 1, 1, 1, 1, 1, 0, 0], 'channel_mult': [1, 2, 4, 4], 'transformer_depth_middle': 1, 'use_linear_in_transformer': True,
            'context_dim': 1024, 'num_head_channels': 64, 'transformer_depth_output': [1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0]}
246
247

    SD21_uncliph = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False,
248
                    'num_classes': 'sequential', 'adm_in_channels': 2048, 'dtype': dtype, 'in_channels': 4, 'model_channels': 320,
249
250
                    'num_res_blocks': [2, 2, 2, 2], 'transformer_depth': [1, 1, 1, 1, 1, 1, 0, 0], 'channel_mult': [1, 2, 4, 4], 'transformer_depth_middle': 1,
                    'use_linear_in_transformer': True, 'context_dim': 1024, 'num_head_channels': 64, 'transformer_depth_output': [1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0]}
251
252

    SD21_unclipl = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False,
253
                    'num_classes': 'sequential', 'adm_in_channels': 1536, 'dtype': dtype, 'in_channels': 4, 'model_channels': 320,
254
255
                    'num_res_blocks': [2, 2, 2, 2], 'transformer_depth': [1, 1, 1, 1, 1, 1, 0, 0], 'channel_mult': [1, 2, 4, 4], 'transformer_depth_middle': 1,
                    'use_linear_in_transformer': True, 'context_dim': 1024, 'num_head_channels': 64, 'transformer_depth_output': [1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0]}
256

257
258
259
260
    SD15 = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False, 'adm_in_channels': None,
            'dtype': dtype, 'in_channels': 4, 'model_channels': 320, 'num_res_blocks': [2, 2, 2, 2], 'transformer_depth': [1, 1, 1, 1, 1, 1, 0, 0],
            'channel_mult': [1, 2, 4, 4], 'transformer_depth_middle': 1, 'use_linear_in_transformer': False, 'context_dim': 768, 'num_heads': 8,
            'transformer_depth_output': [1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0]}
261

262
    SDXL_mid_cnet = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False,
263
264
265
                     'num_classes': 'sequential', 'adm_in_channels': 2816, 'dtype': dtype, 'in_channels': 4, 'model_channels': 320,
                     'num_res_blocks': [2, 2, 2], 'transformer_depth': [0, 0, 0, 0, 1, 1], 'channel_mult': [1, 2, 4], 'transformer_depth_middle': 1,
                     'use_linear_in_transformer': True, 'context_dim': 2048, 'num_head_channels': 64, 'transformer_depth_output': [0, 0, 0, 0, 0, 0, 1, 1, 1]}
266

267
    SDXL_small_cnet = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False,
268
269
270
                       'num_classes': 'sequential', 'adm_in_channels': 2816, 'dtype': dtype, 'in_channels': 4, 'model_channels': 320,
                       'num_res_blocks': [2, 2, 2], 'transformer_depth': [0, 0, 0, 0, 0, 0], 'channel_mult': [1, 2, 4], 'transformer_depth_middle': 0,
                       'use_linear_in_transformer': True, 'num_head_channels': 64, 'context_dim': 1, 'transformer_depth_output': [0, 0, 0, 0, 0, 0, 0, 0, 0]}
271

comfyanonymous's avatar
comfyanonymous committed
272
    SDXL_diffusers_inpaint = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False,
273
274
275
276
277
278
279
280
                              'num_classes': 'sequential', 'adm_in_channels': 2816, 'dtype': dtype, 'in_channels': 9, 'model_channels': 320,
                              'num_res_blocks': [2, 2, 2], 'transformer_depth': [0, 0, 2, 2, 10, 10], 'channel_mult': [1, 2, 4], 'transformer_depth_middle': 10,
                              'use_linear_in_transformer': True, 'context_dim': 2048, 'num_head_channels': 64, 'transformer_depth_output': [0, 0, 0, 2, 2, 2, 10, 10, 10]}

    SSD_1B = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False,
              'num_classes': 'sequential', 'adm_in_channels': 2816, 'dtype': dtype, 'in_channels': 4, 'model_channels': 320,
              'num_res_blocks': [2, 2, 2], 'transformer_depth': [0, 0, 2, 2, 4, 4], 'transformer_depth_output': [0, 0, 0, 1, 1, 2, 10, 4, 4],
              'channel_mult': [1, 2, 4], 'transformer_depth_middle': -1, 'use_linear_in_transformer': True, 'context_dim': 2048, 'num_head_channels': 64}
281

282
    supported_models = [SDXL, SDXL_refiner, SD21, SD15, SD21_uncliph, SD21_unclipl, SDXL_mid_cnet, SDXL_small_cnet, SDXL_diffusers_inpaint, SSD_1B]
283
284
285
286
287
288
289
290

    for unet_config in supported_models:
        matches = True
        for k in match:
            if match[k] != unet_config[k]:
                matches = False
                break
        if matches:
291
            return convert_config(unet_config)
292
293
    return None

294
295
def model_config_from_diffusers_unet(state_dict, dtype):
    unet_config = unet_config_from_diffusers_unet(state_dict, dtype)
296
297
    if unet_config is not None:
        return model_config_from_unet_config(unet_config)
298
    return None