samplers.py 29.9 KB
Newer Older
1
2
from .k_diffusion import sampling as k_diffusion_sampling
from .k_diffusion import external as k_diffusion_external
3
from .extra_samplers import uni_pc
comfyanonymous's avatar
comfyanonymous committed
4
import torch
5
from comfy import model_management
comfyanonymous's avatar
comfyanonymous committed
6
7
from .ldm.models.diffusion.ddim import DDIMSampler
from .ldm.modules.diffusionmodules.util import make_ddim_timesteps
comfyanonymous's avatar
comfyanonymous committed
8
import math
9
from comfy import model_base
comfyanonymous's avatar
comfyanonymous committed
10
11
12

def lcm(a, b): #TODO: eventually replace by math.lcm (added in python3.9)
    return abs(a*b) // math.gcd(a, b)
comfyanonymous's avatar
comfyanonymous committed
13

comfyanonymous's avatar
comfyanonymous committed
14
15
#The main sampling function shared by all the samplers
#Returns predicted noise
16
def sampling_function(model_function, x, timestep, uncond, cond, cond_scale, cond_concat=None, model_options={}, seed=None):
comfyanonymous's avatar
comfyanonymous committed
17
        def get_area_and_mult(cond, x_in, cond_concat_in, timestep_in):
18
19
            area = (x_in.shape[2], x_in.shape[3], 0, 0)
            strength = 1.0
20
21
22
23
24
25
26
27
            if 'timestep_start' in cond[1]:
                timestep_start = cond[1]['timestep_start']
                if timestep_in > timestep_start:
                    return None
            if 'timestep_end' in cond[1]:
                timestep_end = cond[1]['timestep_end']
                if timestep_in < timestep_end:
                    return None
28
29
30
31
            if 'area' in cond[1]:
                area = cond[1]['area']
            if 'strength' in cond[1]:
                strength = cond[1]['strength']
32

33
            adm_cond = None
34
35
            if 'adm_encoded' in cond[1]:
                adm_cond = cond[1]['adm_encoded']
36

37
            input_x = x_in[:,:,area[2]:area[0] + area[2],area[3]:area[1] + area[3]]
Jacob Segal's avatar
Jacob Segal committed
38
39
40
            if 'mask' in cond[1]:
                # Scale the mask to the size of the input
                # The mask should have been resized as we began the sampling process
41
42
43
                mask_strength = 1.0
                if "mask_strength" in cond[1]:
                    mask_strength = cond[1]["mask_strength"]
Jacob Segal's avatar
Jacob Segal committed
44
45
46
                mask = cond[1]['mask']
                assert(mask.shape[1] == x_in.shape[2])
                assert(mask.shape[2] == x_in.shape[3])
47
                mask = mask[:,area[2]:area[0] + area[2],area[3]:area[1] + area[3]] * mask_strength
Jacob Segal's avatar
Jacob Segal committed
48
                mask = mask.unsqueeze(1).repeat(input_x.shape[0] // mask.shape[0], input_x.shape[1], 1, 1)
Jacob Segal's avatar
Jacob Segal committed
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
            else:
                mask = torch.ones_like(input_x)
            mult = mask * strength

            if 'mask' not in cond[1]:
                rr = 8
                if area[2] != 0:
                    for t in range(rr):
                        mult[:,:,t:1+t,:] *= ((1.0/rr) * (t + 1))
                if (area[0] + area[2]) < x_in.shape[2]:
                    for t in range(rr):
                        mult[:,:,area[0] - 1 - t:area[0] - t,:] *= ((1.0/rr) * (t + 1))
                if area[3] != 0:
                    for t in range(rr):
                        mult[:,:,:,t:1+t] *= ((1.0/rr) * (t + 1))
                if (area[1] + area[3]) < x_in.shape[3]:
                    for t in range(rr):
                        mult[:,:,:,area[1] - 1 - t:area[1] - t] *= ((1.0/rr) * (t + 1))

comfyanonymous's avatar
comfyanonymous committed
68
69
70
71
72
73
74
75
            conditionning = {}
            conditionning['c_crossattn'] = cond[0]
            if cond_concat_in is not None and len(cond_concat_in) > 0:
                cropped = []
                for x in cond_concat_in:
                    cr = x[:,:,area[2]:area[0] + area[2],area[3]:area[1] + area[3]]
                    cropped.append(cr)
                conditionning['c_concat'] = torch.cat(cropped, dim=1)
comfyanonymous's avatar
comfyanonymous committed
76

77
78
79
            if adm_cond is not None:
                conditionning['c_adm'] = adm_cond

comfyanonymous's avatar
comfyanonymous committed
80
81
82
            control = None
            if 'control' in cond[1]:
                control = cond[1]['control']
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97

            patches = None
            if 'gligen' in cond[1]:
                gligen = cond[1]['gligen']
                patches = {}
                gligen_type = gligen[0]
                gligen_model = gligen[1]
                if gligen_type == "position":
                    gligen_patch = gligen_model.set_position(input_x.shape, gligen[2], input_x.device)
                else:
                    gligen_patch = gligen_model.set_empty(input_x.shape, input_x.device)

                patches['middle_patch'] = [gligen_patch]

            return (input_x, mult, conditionning, area, control, patches)
comfyanonymous's avatar
comfyanonymous committed
98
99

        def cond_equal_size(c1, c2):
comfyanonymous's avatar
comfyanonymous committed
100
101
            if c1 is c2:
                return True
comfyanonymous's avatar
comfyanonymous committed
102
103
104
            if c1.keys() != c2.keys():
                return False
            if 'c_crossattn' in c1:
comfyanonymous's avatar
comfyanonymous committed
105
106
107
108
109
110
111
112
113
114
                s1 = c1['c_crossattn'].shape
                s2 = c2['c_crossattn'].shape
                if s1 != s2:
                    if s1[0] != s2[0] or s1[2] != s2[2]: #these 2 cases should not happen
                        return False

                    mult_min = lcm(s1[1], s2[1])
                    diff = mult_min // min(s1[1], s2[1])
                    if diff > 4: #arbitrary limit on the padding because it's probably going to impact performance negatively if it's too much
                        return False
comfyanonymous's avatar
comfyanonymous committed
115
116
117
            if 'c_concat' in c1:
                if c1['c_concat'].shape != c2['c_concat'].shape:
                    return False
118
119
120
            if 'c_adm' in c1:
                if c1['c_adm'].shape != c2['c_adm'].shape:
                    return False
comfyanonymous's avatar
comfyanonymous committed
121
122
            return True

comfyanonymous's avatar
comfyanonymous committed
123
124
125
        def can_concat_cond(c1, c2):
            if c1[0].shape != c2[0].shape:
                return False
126
127

            #control
comfyanonymous's avatar
comfyanonymous committed
128
129
130
131
132
133
            if (c1[4] is None) != (c2[4] is None):
                return False
            if c1[4] is not None:
                if c1[4] is not c2[4]:
                    return False

134
135
136
137
138
139
140
            #patches
            if (c1[5] is None) != (c2[5] is None):
                return False
            if (c1[5] is not None):
                if c1[5] is not c2[5]:
                    return False

comfyanonymous's avatar
comfyanonymous committed
141
142
            return cond_equal_size(c1[2], c2[2])

comfyanonymous's avatar
comfyanonymous committed
143
144
145
        def cond_cat(c_list):
            c_crossattn = []
            c_concat = []
146
            c_adm = []
comfyanonymous's avatar
comfyanonymous committed
147
            crossattn_max_len = 0
comfyanonymous's avatar
comfyanonymous committed
148
149
            for x in c_list:
                if 'c_crossattn' in x:
comfyanonymous's avatar
comfyanonymous committed
150
151
152
153
154
155
                    c = x['c_crossattn']
                    if crossattn_max_len == 0:
                        crossattn_max_len = c.shape[1]
                    else:
                        crossattn_max_len = lcm(crossattn_max_len, c.shape[1])
                    c_crossattn.append(c)
comfyanonymous's avatar
comfyanonymous committed
156
157
                if 'c_concat' in x:
                    c_concat.append(x['c_concat'])
158
159
                if 'c_adm' in x:
                    c_adm.append(x['c_adm'])
comfyanonymous's avatar
comfyanonymous committed
160
            out = {}
comfyanonymous's avatar
comfyanonymous committed
161
162
163
164
165
166
167
168
            c_crossattn_out = []
            for c in c_crossattn:
                if c.shape[1] < crossattn_max_len:
                    c = c.repeat(1, crossattn_max_len // c.shape[1], 1) #padding with repeat doesn't change result
                c_crossattn_out.append(c)

            if len(c_crossattn_out) > 0:
                out['c_crossattn'] = [torch.cat(c_crossattn_out)]
comfyanonymous's avatar
comfyanonymous committed
169
170
            if len(c_concat) > 0:
                out['c_concat'] = [torch.cat(c_concat)]
171
172
            if len(c_adm) > 0:
                out['c_adm'] = torch.cat(c_adm)
comfyanonymous's avatar
comfyanonymous committed
173
174
            return out

175
        def calc_cond_uncond_batch(model_function, cond, uncond, x_in, timestep, max_total_area, cond_concat_in, model_options):
comfyanonymous's avatar
comfyanonymous committed
176
177
            out_cond = torch.zeros_like(x_in)
            out_count = torch.ones_like(x_in)/100000.0
178
179
180
181
182
183

            out_uncond = torch.zeros_like(x_in)
            out_uncond_count = torch.ones_like(x_in)/100000.0

            COND = 0
            UNCOND = 1
comfyanonymous's avatar
comfyanonymous committed
184

185
            to_run = []
comfyanonymous's avatar
comfyanonymous committed
186
            for x in cond:
comfyanonymous's avatar
comfyanonymous committed
187
                p = get_area_and_mult(x, x_in, cond_concat_in, timestep)
188
                if p is None:
comfyanonymous's avatar
comfyanonymous committed
189
                    continue
190
191
192

                to_run += [(p, COND)]
            for x in uncond:
comfyanonymous's avatar
comfyanonymous committed
193
                p = get_area_and_mult(x, x_in, cond_concat_in, timestep)
194
195
196
197
198
199
200
201
                if p is None:
                    continue

                to_run += [(p, UNCOND)]

            while len(to_run) > 0:
                first = to_run[0]
                first_shape = first[0][0].shape
202
                to_batch_temp = []
203
                for x in range(len(to_run)):
comfyanonymous's avatar
comfyanonymous committed
204
205
                    if can_concat_cond(to_run[x][0], first[0]):
                        to_batch_temp += [x]
206
207
208
209
210
211
212
213
214

                to_batch_temp.reverse()
                to_batch = to_batch_temp[:1]

                for i in range(1, len(to_batch_temp) + 1):
                    batch_amount = to_batch_temp[:len(to_batch_temp)//i]
                    if (len(batch_amount) * first_shape[0] * first_shape[2] * first_shape[3] < max_total_area):
                        to_batch = batch_amount
                        break
215
216
217
218
219
220

                input_x = []
                mult = []
                c = []
                cond_or_uncond = []
                area = []
comfyanonymous's avatar
comfyanonymous committed
221
                control = None
222
                patches = None
223
224
225
226
227
228
229
230
                for x in to_batch:
                    o = to_run.pop(x)
                    p = o[0]
                    input_x += [p[0]]
                    mult += [p[1]]
                    c += [p[2]]
                    area += [p[3]]
                    cond_or_uncond += [o[1]]
comfyanonymous's avatar
comfyanonymous committed
231
                    control = p[4]
232
                    patches = p[5]
233
234
235

                batch_chunks = len(cond_or_uncond)
                input_x = torch.cat(input_x)
comfyanonymous's avatar
comfyanonymous committed
236
                c = cond_cat(c)
comfyanonymous's avatar
comfyanonymous committed
237
                timestep_ = torch.cat([timestep] * batch_chunks)
238

comfyanonymous's avatar
comfyanonymous committed
239
                if control is not None:
240
                    c['control'] = control.get_control(input_x, timestep_, c, len(cond_or_uncond))
comfyanonymous's avatar
comfyanonymous committed
241

242
                transformer_options = {}
243
                if 'transformer_options' in model_options:
244
245
246
                    transformer_options = model_options['transformer_options'].copy()

                if patches is not None:
247
248
249
250
251
252
253
254
255
                    if "patches" in transformer_options:
                        cur_patches = transformer_options["patches"].copy()
                        for p in patches:
                            if p in cur_patches:
                                cur_patches[p] = cur_patches[p] + patches[p]
                            else:
                                cur_patches[p] = patches[p]
                    else:
                        transformer_options["patches"] = patches
256
257

                c['transformer_options'] = transformer_options
258

259
260
261
262
                if 'model_function_wrapper' in model_options:
                    output = model_options['model_function_wrapper'](model_function, {"input": input_x, "timestep": timestep_, "c": c, "cond_or_uncond": cond_or_uncond}).chunk(batch_chunks)
                else:
                    output = model_function(input_x, timestep_, **c).chunk(batch_chunks)
comfyanonymous's avatar
comfyanonymous committed
263
                del input_x
264

265
266
                model_management.throw_exception_if_processing_interrupted()

267
268
269
270
271
272
273
                for o in range(batch_chunks):
                    if cond_or_uncond[o] == COND:
                        out_cond[:,:,area[o][2]:area[o][0] + area[o][2],area[o][3]:area[o][1] + area[o][3]] += output[o] * mult[o]
                        out_count[:,:,area[o][2]:area[o][0] + area[o][2],area[o][3]:area[o][1] + area[o][3]] += mult[o]
                    else:
                        out_uncond[:,:,area[o][2]:area[o][0] + area[o][2],area[o][3]:area[o][1] + area[o][3]] += output[o] * mult[o]
                        out_uncond_count[:,:,area[o][2]:area[o][0] + area[o][2],area[o][3]:area[o][1] + area[o][3]] += mult[o]
comfyanonymous's avatar
comfyanonymous committed
274
275
276
277
                del mult

            out_cond /= out_count
            del out_count
278
279
280
281
            out_uncond /= out_uncond_count
            del out_uncond_count

            return out_cond, out_uncond
comfyanonymous's avatar
comfyanonymous committed
282
283


284
        max_total_area = model_management.maximum_batch_area()
285
        cond, uncond = calc_cond_uncond_batch(model_function, cond, uncond, x, timestep, max_total_area, cond_concat, model_options)
286
        if "sampler_cfg_function" in model_options:
287
288
            args = {"cond": cond, "uncond": uncond, "cond_scale": cond_scale, "timestep": timestep}
            return model_options["sampler_cfg_function"](args)
289
290
        else:
            return uncond + (cond - uncond) * cond_scale
comfyanonymous's avatar
comfyanonymous committed
291

comfyanonymous's avatar
comfyanonymous committed
292
293
294
295
296
297
298
299
300
301
302
303
304
305

class CompVisVDenoiser(k_diffusion_external.DiscreteVDDPMDenoiser):
    def __init__(self, model, quantize=False, device='cpu'):
        super().__init__(model, model.alphas_cumprod, quantize=quantize)

    def get_v(self, x, t, cond, **kwargs):
        return self.inner_model.apply_model(x, t, cond, **kwargs)


class CFGNoisePredictor(torch.nn.Module):
    def __init__(self, model):
        super().__init__()
        self.inner_model = model
        self.alphas_cumprod = model.alphas_cumprod
306
307
    def apply_model(self, x, timestep, cond, uncond, cond_scale, cond_concat=None, model_options={}, seed=None):
        out = sampling_function(self.inner_model.apply_model, x, timestep, uncond, cond, cond_scale, cond_concat, model_options=model_options, seed=seed)
comfyanonymous's avatar
comfyanonymous committed
308
309
310
311
        return out


class KSamplerX0Inpaint(torch.nn.Module):
312
313
314
    def __init__(self, model):
        super().__init__()
        self.inner_model = model
315
    def forward(self, x, sigma, uncond, cond, cond_scale, denoise_mask, cond_concat=None, model_options={}, seed=None):
316
317
        if denoise_mask is not None:
            latent_mask = 1. - denoise_mask
318
            x = x * denoise_mask + (self.latent_image + self.noise * sigma.reshape([sigma.shape[0]] + [1] * (len(self.noise.shape) - 1))) * latent_mask
319
        out = self.inner_model(x, sigma, cond=cond, uncond=uncond, cond_scale=cond_scale, cond_concat=cond_concat, model_options=model_options, seed=seed)
320
321
322
323
324
325
        if denoise_mask is not None:
            out *= denoise_mask

        if denoise_mask is not None:
            out += self.latent_image * latent_mask
        return out
326

comfyanonymous's avatar
comfyanonymous committed
327
328
329
330
331
332
333
334
def simple_scheduler(model, steps):
    sigs = []
    ss = len(model.sigmas) / steps
    for x in range(steps):
        sigs += [float(model.sigmas[-(1 + int(x * ss))])]
    sigs += [0.0]
    return torch.FloatTensor(sigs)

comfyanonymous's avatar
comfyanonymous committed
335
336
337
338
def ddim_scheduler(model, steps):
    sigs = []
    ddim_timesteps = make_ddim_timesteps(ddim_discr_method="uniform", num_ddim_timesteps=steps, num_ddpm_timesteps=model.inner_model.inner_model.num_timesteps, verbose=False)
    for x in range(len(ddim_timesteps) - 1, -1, -1):
339
340
341
342
        ts = ddim_timesteps[x]
        if ts > 999:
            ts = 999
        sigs.append(model.t_to_sigma(torch.tensor(ts)))
comfyanonymous's avatar
comfyanonymous committed
343
344
345
    sigs += [0.0]
    return torch.FloatTensor(sigs)

comfyanonymous's avatar
comfyanonymous committed
346
347
348
349
350
351
352
353
354
def blank_inpaint_image_like(latent_image):
    blank_image = torch.ones_like(latent_image)
    # these are the values for "zero" in pixel space translated to latent space
    blank_image[:,0] *= 0.8223
    blank_image[:,1] *= -0.6876
    blank_image[:,2] *= 0.6364
    blank_image[:,3] *= 0.1380
    return blank_image

Jacob Segal's avatar
Jacob Segal committed
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
def get_mask_aabb(masks):
    if masks.numel() == 0:
        return torch.zeros((0, 4), device=masks.device, dtype=torch.int)

    b = masks.shape[0]

    bounding_boxes = torch.zeros((b, 4), device=masks.device, dtype=torch.int)
    is_empty = torch.zeros((b), device=masks.device, dtype=torch.bool)
    for i in range(b):
        mask = masks[i]
        if mask.numel() == 0:
            continue
        if torch.max(mask != 0) == False:
            is_empty[i] = True
            continue
        y, x = torch.where(mask)
        bounding_boxes[i, 0] = torch.min(x)
        bounding_boxes[i, 1] = torch.min(y)
        bounding_boxes[i, 2] = torch.max(x)
        bounding_boxes[i, 3] = torch.max(y)

    return bounding_boxes, is_empty

Jacob Segal's avatar
Jacob Segal committed
378
379
380
381
382
383
384
385
386
387
388
def resolve_cond_masks(conditions, h, w, device):
    # We need to decide on an area outside the sampling loop in order to properly generate opposite areas of equal sizes.
    # While we're doing this, we can also resolve the mask device and scaling for performance reasons
    for i in range(len(conditions)):
        c = conditions[i]
        if 'mask' in c[1]:
            mask = c[1]['mask']
            mask = mask.to(device=device)
            modified = c[1].copy()
            if len(mask.shape) == 2:
                mask = mask.unsqueeze(0)
mara's avatar
mara committed
389
            if mask.shape[1] != h or mask.shape[2] != w:
Jacob Segal's avatar
Jacob Segal committed
390
391
                mask = torch.nn.functional.interpolate(mask.unsqueeze(1), size=(h, w), mode='bilinear', align_corners=False).squeeze(1)

Jacob Segal's avatar
Jacob Segal committed
392
            if modified.get("set_area_to_bounds", False):
Jacob Segal's avatar
Jacob Segal committed
393
                bounds = torch.max(torch.abs(mask),dim=0).values.unsqueeze(0)
Jacob Segal's avatar
Jacob Segal committed
394
395
396
397
                boxes, is_empty = get_mask_aabb(bounds)
                if is_empty[0]:
                    # Use the minimum possible size for efficiency reasons. (Since the mask is all-0, this becomes a noop anyway)
                    modified['area'] = (8, 8, 0, 0)
Jacob Segal's avatar
Jacob Segal committed
398
                else:
Jacob Segal's avatar
Jacob Segal committed
399
                    box = boxes[0]
Jacob Segal's avatar
Jacob Segal committed
400
                    H, W, Y, X = (box[3] - box[1] + 1, box[2] - box[0] + 1, box[1], box[0])
401
402
                    H = max(8, H)
                    W = max(8, W)
Jacob Segal's avatar
Jacob Segal committed
403
404
                    area = (int(H), int(W), int(Y), int(X))
                    modified['area'] = area
Jacob Segal's avatar
Jacob Segal committed
405
406
407
408

            modified['mask'] = mask
            conditions[i] = [c[0], modified]

comfyanonymous's avatar
comfyanonymous committed
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
def create_cond_with_same_area_if_none(conds, c):
    if 'area' not in c[1]:
        return

    c_area = c[1]['area']
    smallest = None
    for x in conds:
        if 'area' in x[1]:
            a = x[1]['area']
            if c_area[2] >= a[2] and c_area[3] >= a[3]:
                if a[0] + a[2] >= c_area[0] + c_area[2]:
                    if a[1] + a[3] >= c_area[1] + c_area[3]:
                        if smallest is None:
                            smallest = x
                        elif 'area' not in smallest[1]:
                            smallest = x
                        else:
                            if smallest[1]['area'][0] * smallest[1]['area'][1] > a[0] * a[1]:
                                smallest = x
        else:
            if smallest is None:
                smallest = x
    if smallest is None:
        return
    if 'area' in smallest[1]:
        if smallest[1]['area'] == c_area:
            return
    n = c[1].copy()
    conds += [[smallest[0], n]]
comfyanonymous's avatar
comfyanonymous committed
438

439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
def calculate_start_end_timesteps(model, conds):
    for t in range(len(conds)):
        x = conds[t]

        timestep_start = None
        timestep_end = None
        if 'start_percent' in x[1]:
            timestep_start = model.sigma_to_t(model.t_to_sigma(torch.tensor(x[1]['start_percent'] * 999.0)))
        if 'end_percent' in x[1]:
            timestep_end = model.sigma_to_t(model.t_to_sigma(torch.tensor(x[1]['end_percent'] * 999.0)))

        if (timestep_start is not None) or (timestep_end is not None):
            n = x[1].copy()
            if (timestep_start is not None):
                n['timestep_start'] = timestep_start
            if (timestep_end is not None):
                n['timestep_end'] = timestep_end
            conds[t] = [x[0], n]

458
def apply_empty_x_to_equal_area(conds, uncond, name, uncond_fill_func):
comfyanonymous's avatar
comfyanonymous committed
459
460
461
462
463
464
465
    cond_cnets = []
    cond_other = []
    uncond_cnets = []
    uncond_other = []
    for t in range(len(conds)):
        x = conds[t]
        if 'area' not in x[1]:
466
467
            if name in x[1] and x[1][name] is not None:
                cond_cnets.append(x[1][name])
comfyanonymous's avatar
comfyanonymous committed
468
469
470
471
472
            else:
                cond_other.append((x, t))
    for t in range(len(uncond)):
        x = uncond[t]
        if 'area' not in x[1]:
473
474
            if name in x[1] and x[1][name] is not None:
                uncond_cnets.append(x[1][name])
comfyanonymous's avatar
comfyanonymous committed
475
476
477
478
479
480
481
482
483
            else:
                uncond_other.append((x, t))

    if len(uncond_cnets) > 0:
        return

    for x in range(len(cond_cnets)):
        temp = uncond_other[x % len(uncond_other)]
        o = temp[0]
484
        if name in o[1] and o[1][name] is not None:
comfyanonymous's avatar
comfyanonymous committed
485
            n = o[1].copy()
486
            n[name] = uncond_fill_func(cond_cnets, x)
comfyanonymous's avatar
comfyanonymous committed
487
488
489
            uncond += [[o[0], n]]
        else:
            n = o[1].copy()
490
            n[name] = uncond_fill_func(cond_cnets, x)
comfyanonymous's avatar
comfyanonymous committed
491
492
            uncond[temp[1]] = [o[0], n]

493
def encode_adm(model, conds, batch_size, width, height, device, prompt_type):
494
495
    for t in range(len(conds)):
        x = conds[t]
comfyanonymous's avatar
comfyanonymous committed
496
        adm_out = None
497
498
        if 'adm' in x[1]:
            adm_out = x[1]["adm"]
499
        else:
500
            params = x[1].copy()
501
502
503
            params["width"] = params.get("width", width * 8)
            params["height"] = params.get("height", height * 8)
            params["prompt_type"] = params.get("prompt_type", prompt_type)
504
            adm_out = model.encode_adm(device=device, **params)
505

comfyanonymous's avatar
comfyanonymous committed
506
507
        if adm_out is not None:
            x[1] = x[1].copy()
508
            x[1]["adm_encoded"] = torch.cat([adm_out] * batch_size).to(device)
509
510
511

    return conds

512

comfyanonymous's avatar
comfyanonymous committed
513
class KSampler:
514
    SCHEDULERS = ["normal", "karras", "exponential", "simple", "ddim_uniform"]
515
    SAMPLERS = ["euler", "euler_ancestral", "heun", "dpm_2", "dpm_2_ancestral",
516
517
                "lms", "dpm_fast", "dpm_adaptive", "dpmpp_2s_ancestral", "dpmpp_sde", "dpmpp_sde_gpu",
                "dpmpp_2m", "dpmpp_2m_sde", "dpmpp_2m_sde_gpu", "ddim", "uni_pc", "uni_pc_bh2"]
comfyanonymous's avatar
comfyanonymous committed
518

519
    def __init__(self, model, steps, device, sampler=None, scheduler=None, denoise=None, model_options={}):
comfyanonymous's avatar
comfyanonymous committed
520
        self.model = model
comfyanonymous's avatar
comfyanonymous committed
521
        self.model_denoise = CFGNoisePredictor(self.model)
522
        if self.model.model_type == model_base.ModelType.V_PREDICTION:
comfyanonymous's avatar
comfyanonymous committed
523
            self.model_wrap = CompVisVDenoiser(self.model_denoise, quantize=True)
comfyanonymous's avatar
comfyanonymous committed
524
        else:
comfyanonymous's avatar
comfyanonymous committed
525
            self.model_wrap = k_diffusion_external.CompVisDenoiser(self.model_denoise, quantize=True)
526

comfyanonymous's avatar
comfyanonymous committed
527
        self.model_k = KSamplerX0Inpaint(self.model_wrap)
comfyanonymous's avatar
comfyanonymous committed
528
529
530
531
532
533
534
        self.device = device
        if scheduler not in self.SCHEDULERS:
            scheduler = self.SCHEDULERS[0]
        if sampler not in self.SAMPLERS:
            sampler = self.SAMPLERS[0]
        self.scheduler = scheduler
        self.sampler = sampler
535
536
        self.sigma_min=float(self.model_wrap.sigma_min)
        self.sigma_max=float(self.model_wrap.sigma_max)
comfyanonymous's avatar
comfyanonymous committed
537
        self.set_steps(steps, denoise)
538
        self.denoise = denoise
539
        self.model_options = model_options
comfyanonymous's avatar
comfyanonymous committed
540

comfyanonymous's avatar
comfyanonymous committed
541
542
543
544
545
546
547
548
549
550
    def calculate_sigmas(self, steps):
        sigmas = None

        discard_penultimate_sigma = False
        if self.sampler in ['dpm_2', 'dpm_2_ancestral']:
            steps += 1
            discard_penultimate_sigma = True

        if self.scheduler == "karras":
            sigmas = k_diffusion_sampling.get_sigmas_karras(n=steps, sigma_min=self.sigma_min, sigma_max=self.sigma_max)
551
552
        elif self.scheduler == "exponential":
            sigmas = k_diffusion_sampling.get_sigmas_exponential(n=steps, sigma_min=self.sigma_min, sigma_max=self.sigma_max)
comfyanonymous's avatar
comfyanonymous committed
553
554
555
556
557
558
559
560
561
562
563
564
565
        elif self.scheduler == "normal":
            sigmas = self.model_wrap.get_sigmas(steps)
        elif self.scheduler == "simple":
            sigmas = simple_scheduler(self.model_wrap, steps)
        elif self.scheduler == "ddim_uniform":
            sigmas = ddim_scheduler(self.model_wrap, steps)
        else:
            print("error invalid scheduler", self.scheduler)

        if discard_penultimate_sigma:
            sigmas = torch.cat([sigmas[:-2], sigmas[-1:]])
        return sigmas

comfyanonymous's avatar
comfyanonymous committed
566
567
    def set_steps(self, steps, denoise=None):
        self.steps = steps
568
        if denoise is None or denoise > 0.9999:
comfyanonymous's avatar
comfyanonymous committed
569
            self.sigmas = self.calculate_sigmas(steps).to(self.device)
comfyanonymous's avatar
comfyanonymous committed
570
571
        else:
            new_steps = int(steps/denoise)
comfyanonymous's avatar
comfyanonymous committed
572
            sigmas = self.calculate_sigmas(new_steps).to(self.device)
comfyanonymous's avatar
comfyanonymous committed
573
574
            self.sigmas = sigmas[-(steps + 1):]

575
    def sample(self, noise, positive, negative, cfg, latent_image=None, start_step=None, last_step=None, force_full_denoise=False, denoise_mask=None, sigmas=None, callback=None, disable_pbar=False, seed=None):
576
577
        if sigmas is None:
            sigmas = self.sigmas
comfyanonymous's avatar
comfyanonymous committed
578
579
        sigma_min = self.sigma_min

comfyanonymous's avatar
comfyanonymous committed
580
        if last_step is not None and last_step < (len(sigmas) - 1):
comfyanonymous's avatar
comfyanonymous committed
581
582
            sigma_min = sigmas[last_step]
            sigmas = sigmas[:last_step + 1]
comfyanonymous's avatar
comfyanonymous committed
583
584
585
            if force_full_denoise:
                sigmas[-1] = 0

comfyanonymous's avatar
comfyanonymous committed
586
        if start_step is not None:
comfyanonymous's avatar
comfyanonymous committed
587
588
589
590
591
592
593
            if start_step < (len(sigmas) - 1):
                sigmas = sigmas[start_step:]
            else:
                if latent_image is not None:
                    return latent_image
                else:
                    return torch.zeros_like(noise)
comfyanonymous's avatar
comfyanonymous committed
594

comfyanonymous's avatar
comfyanonymous committed
595
596
        positive = positive[:]
        negative = negative[:]
Jacob Segal's avatar
Jacob Segal committed
597
598
599
600

        resolve_cond_masks(positive, noise.shape[2], noise.shape[3], self.device)
        resolve_cond_masks(negative, noise.shape[2], noise.shape[3], self.device)

601
602
603
        calculate_start_end_timesteps(self.model_wrap, negative)
        calculate_start_end_timesteps(self.model_wrap, positive)

comfyanonymous's avatar
comfyanonymous committed
604
605
606
607
608
609
        #make sure each cond area has an opposite one with the same area
        for c in positive:
            create_cond_with_same_area_if_none(negative, c)
        for c in negative:
            create_cond_with_same_area_if_none(positive, c)

610
611
        apply_empty_x_to_equal_area(positive, negative, 'control', lambda cond_cnets, x: cond_cnets[x])
        apply_empty_x_to_equal_area(positive, negative, 'gligen', lambda cond_cnets, x: cond_cnets[x])
comfyanonymous's avatar
comfyanonymous committed
612

comfyanonymous's avatar
comfyanonymous committed
613
        if self.model.is_adm():
614
615
            positive = encode_adm(self.model, positive, noise.shape[0], noise.shape[3], noise.shape[2], self.device, "positive")
            negative = encode_adm(self.model, negative, noise.shape[0], noise.shape[3], noise.shape[2], self.device, "negative")
616

617
618
619
        if latent_image is not None:
            latent_image = self.model.process_latent_in(latent_image)

620
        extra_args = {"cond":positive, "uncond":negative, "cond_scale": cfg, "model_options": self.model_options, "seed":seed}
comfyanonymous's avatar
comfyanonymous committed
621

comfyanonymous's avatar
comfyanonymous committed
622
        cond_concat = None
623
        if hasattr(self.model, 'concat_keys'): #inpaint
comfyanonymous's avatar
comfyanonymous committed
624
625
626
627
628
629
            cond_concat = []
            for ck in self.model.concat_keys:
                if denoise_mask is not None:
                    if ck == "mask":
                        cond_concat.append(denoise_mask[:,:1])
                    elif ck == "masked_image":
630
                        cond_concat.append(latent_image) #NOTE: the latent_image should be masked by the mask in pixel space
comfyanonymous's avatar
comfyanonymous committed
631
632
633
634
635
636
637
                else:
                    if ck == "mask":
                        cond_concat.append(torch.ones_like(noise)[:,:1])
                    elif ck == "masked_image":
                        cond_concat.append(blank_inpaint_image_like(noise))
            extra_args["cond_concat"] = cond_concat

638
639
640
641
642
        if sigmas[0] != self.sigmas[0] or (self.denoise is not None and self.denoise < 1.0):
            max_denoise = False
        else:
            max_denoise = True

643

644
645
646
647
648
649
650
        if self.sampler == "uni_pc":
            samples = uni_pc.sample_unipc(self.model_wrap, noise, latent_image, sigmas, sampling_function=sampling_function, max_denoise=max_denoise, extra_args=extra_args, noise_mask=denoise_mask, callback=callback, disable=disable_pbar)
        elif self.sampler == "uni_pc_bh2":
            samples = uni_pc.sample_unipc(self.model_wrap, noise, latent_image, sigmas, sampling_function=sampling_function, max_denoise=max_denoise, extra_args=extra_args, noise_mask=denoise_mask, callback=callback, variant='bh2', disable=disable_pbar)
        elif self.sampler == "ddim":
            timesteps = []
            for s in range(sigmas.shape[0]):
651
                timesteps.insert(0, self.model_wrap.sigma_to_discrete_timestep(sigmas[s]))
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
            noise_mask = None
            if denoise_mask is not None:
                noise_mask = 1.0 - denoise_mask

            ddim_callback = None
            if callback is not None:
                total_steps = len(timesteps) - 1
                ddim_callback = lambda pred_x0, i: callback(i, pred_x0, None, total_steps)

            sampler = DDIMSampler(self.model, device=self.device)
            sampler.make_schedule_timesteps(ddim_timesteps=timesteps, verbose=False)
            z_enc = sampler.stochastic_encode(latent_image, torch.tensor([len(timesteps) - 1] * noise.shape[0]).to(self.device), noise=noise, max_denoise=max_denoise)
            samples, _ = sampler.sample_custom(ddim_timesteps=timesteps,
                                                    conditioning=positive,
                                                    batch_size=noise.shape[0],
                                                    shape=noise.shape[1:],
                                                    verbose=False,
                                                    unconditional_guidance_scale=cfg,
                                                    unconditional_conditioning=negative,
                                                    eta=0.0,
                                                    x_T=z_enc,
                                                    x0=latent_image,
                                                    img_callback=ddim_callback,
675
                                                    denoise_function=self.model_wrap.predict_eps_discrete_timestep,
676
677
678
679
680
                                                    extra_args=extra_args,
                                                    mask=noise_mask,
                                                    to_zero=sigmas[-1]==0,
                                                    end_step=sigmas.shape[0] - 1,
                                                    disable_pbar=disable_pbar)
681

682
683
684
685
        else:
            extra_args["denoise_mask"] = denoise_mask
            self.model_k.latent_image = latent_image
            self.model_k.noise = noise
686

687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
            if max_denoise:
                noise = noise * torch.sqrt(1.0 + sigmas[0] ** 2.0)
            else:
                noise = noise * sigmas[0]

            k_callback = None
            total_steps = len(sigmas) - 1
            if callback is not None:
                k_callback = lambda x: callback(x["i"], x["denoised"], x["x"], total_steps)

            if latent_image is not None:
                noise += latent_image
            if self.sampler == "dpm_fast":
                samples = k_diffusion_sampling.sample_dpm_fast(self.model_k, noise, sigma_min, sigmas[0], total_steps, extra_args=extra_args, callback=k_callback, disable=disable_pbar)
            elif self.sampler == "dpm_adaptive":
                samples = k_diffusion_sampling.sample_dpm_adaptive(self.model_k, noise, sigma_min, sigmas[0], extra_args=extra_args, callback=k_callback, disable=disable_pbar)
            else:
                samples = getattr(k_diffusion_sampling, "sample_{}".format(self.sampler))(self.model_k, noise, sigmas, extra_args=extra_args, callback=k_callback, disable=disable_pbar)
705

706
        return self.model.process_latent_out(samples.to(torch.float32))