File size: 17,099 Bytes
9ab270d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
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
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
203
204
205
206
207
208
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
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
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
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
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
405
406
407
408
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
import numpy as np
import copy

from tqdm.auto import trange
from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl_img2img import *
from diffusers.models.transformers import Transformer2DModel


original_Transformer2DModel_forward = Transformer2DModel.forward


def hacked_Transformer2DModel_forward(
        self,
        hidden_states: torch.Tensor,
        encoder_hidden_states: Optional[torch.Tensor] = None,
        timestep: Optional[torch.LongTensor] = None,
        added_cond_kwargs: Dict[str, torch.Tensor] = None,
        class_labels: Optional[torch.LongTensor] = None,
        cross_attention_kwargs: Dict[str, Any] = None,
        attention_mask: Optional[torch.Tensor] = None,
        encoder_attention_mask: Optional[torch.Tensor] = None,
        return_dict: bool = True,
):
    cross_attention_kwargs = cross_attention_kwargs or {}
    cross_attention_kwargs['hidden_states_original_shape'] = hidden_states.shape
    return original_Transformer2DModel_forward(
        self, hidden_states, encoder_hidden_states, timestep, added_cond_kwargs, class_labels, cross_attention_kwargs,
        attention_mask, encoder_attention_mask, return_dict)


Transformer2DModel.forward = hacked_Transformer2DModel_forward


@torch.no_grad()
def sample_dpmpp_2m(model, x, sigmas, extra_args=None, callback=None, disable=None):
    """DPM-Solver++(2M)."""
    extra_args = {} if extra_args is None else extra_args
    s_in = x.new_ones([x.shape[0]])
    sigma_fn = lambda t: t.neg().exp()
    t_fn = lambda sigma: sigma.log().neg()
    old_denoised = None

    for i in trange(len(sigmas) - 1, disable=disable):
        denoised = model(x, sigmas[i] * s_in, **extra_args)
        if callback is not None:
            callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
        t, t_next = t_fn(sigmas[i]), t_fn(sigmas[i + 1])
        h = t_next - t
        if old_denoised is None or sigmas[i + 1] == 0:
            x = (sigma_fn(t_next) / sigma_fn(t)) * x - (-h).expm1() * denoised
        else:
            h_last = t - t_fn(sigmas[i - 1])
            r = h_last / h
            denoised_d = (1 + 1 / (2 * r)) * denoised - (1 / (2 * r)) * old_denoised
            x = (sigma_fn(t_next) / sigma_fn(t)) * x - (-h).expm1() * denoised_d
        old_denoised = denoised
    return x


class KModel:
    def __init__(self, unet, timesteps=1000, linear_start=0.00085, linear_end=0.012):
        betas = torch.linspace(linear_start ** 0.5, linear_end ** 0.5, timesteps, dtype=torch.float64) ** 2
        alphas = 1. - betas
        alphas_cumprod = torch.tensor(np.cumprod(alphas, axis=0), dtype=torch.float32)

        self.sigmas = ((1 - alphas_cumprod) / alphas_cumprod) ** 0.5
        self.log_sigmas = self.sigmas.log()
        self.sigma_data = 1.0
        self.unet = unet
        return

    @property
    def sigma_min(self):
        return self.sigmas[0]

    @property
    def sigma_max(self):
        return self.sigmas[-1]

    def timestep(self, sigma):
        log_sigma = sigma.log()
        dists = log_sigma.to(self.log_sigmas.device) - self.log_sigmas[:, None]
        return dists.abs().argmin(dim=0).view(sigma.shape).to(sigma.device)

    def get_sigmas_karras(self, n, rho=7.):
        ramp = torch.linspace(0, 1, n)
        min_inv_rho = self.sigma_min ** (1 / rho)
        max_inv_rho = self.sigma_max ** (1 / rho)
        sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho
        return torch.cat([sigmas, sigmas.new_zeros([1])])

    def __call__(self, x, sigma, **extra_args):
        x_ddim_space = x / (sigma[:, None, None, None] ** 2 + self.sigma_data ** 2) ** 0.5
        t = self.timestep(sigma)
        cfg_scale = extra_args['cfg_scale']
        eps_positive = self.unet(x_ddim_space, t, return_dict=False, **extra_args['positive'])[0]
        eps_negative = self.unet(x_ddim_space, t, return_dict=False, **extra_args['negative'])[0]
        noise_pred = eps_negative + cfg_scale * (eps_positive - eps_negative)
        return x - noise_pred * sigma[:, None, None, None]


class OmostSelfAttnProcessor:
    def __call__(self, attn, hidden_states, encoder_hidden_states, hidden_states_original_shape, *args, **kwargs):
        batch_size, sequence_length, _ = hidden_states.shape

        query = attn.to_q(hidden_states)
        key = attn.to_k(hidden_states)
        value = attn.to_v(hidden_states)

        inner_dim = key.shape[-1]
        head_dim = inner_dim // attn.heads

        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)

        hidden_states = torch.nn.functional.scaled_dot_product_attention(
            query, key, value, attn_mask=None, dropout_p=0.0, is_causal=False
        )

        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
        hidden_states = hidden_states.to(query.dtype)
        hidden_states = attn.to_out[0](hidden_states)
        hidden_states = attn.to_out[1](hidden_states)

        return hidden_states


class OmostCrossAttnProcessor:
    def __call__(self, attn, hidden_states, encoder_hidden_states, hidden_states_original_shape, *args, **kwargs):
        B, C, H, W = hidden_states_original_shape

        conds = []
        masks = []

        for m, c in encoder_hidden_states:
            m = torch.nn.functional.interpolate(m[None, None, :, :], (H, W), mode='nearest-exact').flatten().unsqueeze(1).repeat(1, c.size(1))
            conds.append(c)
            masks.append(m)

        conds = torch.cat(conds, dim=1)
        masks = torch.cat(masks, dim=1)

        mask_bool = masks > 0.5
        mask_scale = (H * W) / torch.sum(masks, dim=0, keepdim=True)

        batch_size, sequence_length, _ = conds.shape

        query = attn.to_q(hidden_states)
        key = attn.to_k(conds)
        value = attn.to_v(conds)

        inner_dim = key.shape[-1]
        head_dim = inner_dim // attn.heads

        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)

        mask_bool = mask_bool[None, None, :, :].repeat(query.size(0), query.size(1), 1, 1)
        mask_scale = mask_scale[None, None, :, :].repeat(query.size(0), query.size(1), 1, 1)

        sim = query @ key.transpose(-2, -1) * attn.scale
        sim = sim * mask_scale.to(sim)
        sim.masked_fill_(mask_bool.logical_not(), float("-inf"))
        sim = sim.softmax(dim=-1)

        h = sim @ value
        h = h.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)

        h = attn.to_out[0](h)
        h = attn.to_out[1](h)
        return h


class StableDiffusionXLOmostPipeline(StableDiffusionXLImg2ImgPipeline):
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)
        self.k_model = KModel(unet=self.unet)

        attn_procs = {}
        for name in self.unet.attn_processors.keys():
            if name.endswith("attn2.processor"):
                attn_procs[name] = OmostCrossAttnProcessor()
            else:
                attn_procs[name] = OmostSelfAttnProcessor()

        self.unet.set_attn_processor(attn_procs)
        return

    @torch.inference_mode()
    def encode_bag_of_subprompts_greedy(self, prefixes: list[str], suffixes: list[str]):
        device = self.text_encoder.device

        @torch.inference_mode()
        def greedy_partition(items, max_sum):
            bags = []
            current_bag = []
            current_sum = 0

            for item in items:
                num = item['length']
                if current_sum + num > max_sum:
                    if current_bag:
                        bags.append(current_bag)
                    current_bag = [item]
                    current_sum = num
                else:
                    current_bag.append(item)
                    current_sum += num

            if current_bag:
                bags.append(current_bag)

            return bags

        @torch.inference_mode()
        def get_77_tokens_in_torch(subprompt_inds, tokenizer):
            # Note that all subprompt are theoretically less than 75 tokens (without bos/eos)
            result = [tokenizer.bos_token_id] + subprompt_inds[:75] + [tokenizer.eos_token_id] + [tokenizer.pad_token_id] * 75
            result = result[:77]
            result = torch.tensor([result]).to(device=device, dtype=torch.int64)
            return result

        @torch.inference_mode()
        def merge_with_prefix(bag):
            merged_ids_t1 = copy.deepcopy(prefix_ids_t1)
            merged_ids_t2 = copy.deepcopy(prefix_ids_t2)

            for item in bag:
                merged_ids_t1.extend(item['ids_t1'])
                merged_ids_t2.extend(item['ids_t2'])

            return dict(
                ids_t1=get_77_tokens_in_torch(merged_ids_t1, self.tokenizer),
                ids_t2=get_77_tokens_in_torch(merged_ids_t2, self.tokenizer_2)
            )

        @torch.inference_mode()
        def double_encode(pair_of_inds):
            inds = [pair_of_inds['ids_t1'], pair_of_inds['ids_t2']]
            text_encoders = [self.text_encoder, self.text_encoder_2]

            pooled_prompt_embeds = None
            prompt_embeds_list = []

            for text_input_ids, text_encoder in zip(inds, text_encoders):
                prompt_embeds = text_encoder(text_input_ids, output_hidden_states=True)

                # Only last pooler_output is needed
                pooled_prompt_embeds = prompt_embeds.pooler_output

                # "2" because SDXL always indexes from the penultimate layer.
                prompt_embeds = prompt_embeds.hidden_states[-2]
                prompt_embeds_list.append(prompt_embeds)

            prompt_embeds = torch.concat(prompt_embeds_list, dim=-1)
            return prompt_embeds, pooled_prompt_embeds

        # Begin with tokenizing prefixes

        prefix_length = 0
        prefix_ids_t1 = []
        prefix_ids_t2 = []

        for prefix in prefixes:
            ids_t1 = self.tokenizer(prefix, truncation=False, add_special_tokens=False).input_ids
            ids_t2 = self.tokenizer_2(prefix, truncation=False, add_special_tokens=False).input_ids
            assert len(ids_t1) == len(ids_t2)
            prefix_length += len(ids_t1)
            prefix_ids_t1 += ids_t1
            prefix_ids_t2 += ids_t2

        # Then tokenizing suffixes

        allowed_suffix_length = 75 - prefix_length
        suffix_targets = []

        for subprompt in suffixes:
            # Note that all subprompt are theoretically less than 75 tokens (without bos/eos)
            # So we can safely just crop it to 75
            ids_t1 = self.tokenizer(subprompt, truncation=False, add_special_tokens=False).input_ids[:75]
            ids_t2 = self.tokenizer_2(subprompt, truncation=False, add_special_tokens=False).input_ids[:75]
            assert len(ids_t1) == len(ids_t2)
            suffix_targets.append(dict(
                length=len(ids_t1),
                ids_t1=ids_t1,
                ids_t2=ids_t2
            ))

        # Then merge prefix and suffix tokens

        suffix_targets = greedy_partition(suffix_targets, max_sum=allowed_suffix_length)
        targets = [merge_with_prefix(b) for b in suffix_targets]

        # Encode!

        conds, poolers = [], []

        for target in targets:
            cond, pooler = double_encode(target)
            conds.append(cond)
            poolers.append(pooler)

        conds_merged = torch.concat(conds, dim=1)
        poolers_merged = poolers[0]

        return dict(cond=conds_merged, pooler=poolers_merged)

    @torch.inference_mode()
    def all_conds_from_canvas(self, canvas_outputs, negative_prompt):
        mask_all = torch.ones(size=(90, 90), dtype=torch.float32)
        negative_cond, negative_pooler = self.encode_cropped_prompt_77tokens(negative_prompt)
        negative_result = [(mask_all, negative_cond)]

        positive_result = []
        positive_pooler = None

        for item in canvas_outputs['bag_of_conditions']:
            current_mask = torch.from_numpy(item['mask']).to(torch.float32)
            current_prefixes = item['prefixes']
            current_suffixes = item['suffixes']

            current_cond = self.encode_bag_of_subprompts_greedy(prefixes=current_prefixes, suffixes=current_suffixes)

            if positive_pooler is None:
                positive_pooler = current_cond['pooler']

            positive_result.append((current_mask, current_cond['cond']))

        return positive_result, positive_pooler, negative_result, negative_pooler

    @torch.inference_mode()
    def encode_cropped_prompt_77tokens(self, prompt: str):
        device = self.text_encoder.device
        tokenizers = [self.tokenizer, self.tokenizer_2]
        text_encoders = [self.text_encoder, self.text_encoder_2]

        pooled_prompt_embeds = None
        prompt_embeds_list = []

        for tokenizer, text_encoder in zip(tokenizers, text_encoders):
            text_input_ids = tokenizer(
                prompt,
                padding="max_length",
                max_length=tokenizer.model_max_length,
                truncation=True,
                return_tensors="pt",
            ).input_ids

            prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True)

            # Only last pooler_output is needed
            pooled_prompt_embeds = prompt_embeds.pooler_output

            # "2" because SDXL always indexes from the penultimate layer.
            prompt_embeds = prompt_embeds.hidden_states[-2]
            prompt_embeds_list.append(prompt_embeds)

        prompt_embeds = torch.concat(prompt_embeds_list, dim=-1)
        prompt_embeds = prompt_embeds.to(dtype=self.unet.dtype, device=device)

        return prompt_embeds, pooled_prompt_embeds

    @torch.inference_mode()
    def __call__(
            self,
            initial_latent: torch.FloatTensor = None,
            strength: float = 1.0,
            num_inference_steps: int = 25,
            guidance_scale: float = 5.0,
            batch_size: Optional[int] = 1,
            generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
            prompt_embeds: Optional[torch.FloatTensor] = None,
            negative_prompt_embeds: Optional[torch.FloatTensor] = None,
            pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
            negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
            cross_attention_kwargs: Optional[dict] = None,
    ):

        device = self.unet.device
        cross_attention_kwargs = cross_attention_kwargs or {}

        # Sigmas

        sigmas = self.k_model.get_sigmas_karras(int(num_inference_steps / strength))
        sigmas = sigmas[-(num_inference_steps + 1):].to(device)

        # Initial latents

        _, C, H, W = initial_latent.shape
        noise = randn_tensor((batch_size, C, H, W), generator=generator, device=device, dtype=self.unet.dtype)
        latents = initial_latent.to(noise) + noise * sigmas[0].to(noise)

        # Shape

        height, width = latents.shape[-2:]
        height = height * self.vae_scale_factor
        width = width * self.vae_scale_factor

        add_time_ids = list((height, width) + (0, 0) + (height, width))
        add_time_ids = torch.tensor([add_time_ids], dtype=self.unet.dtype)
        add_neg_time_ids = add_time_ids.clone()

        # Batch

        latents = latents.to(device)
        add_time_ids = add_time_ids.repeat(batch_size, 1).to(device)
        add_neg_time_ids = add_neg_time_ids.repeat(batch_size, 1).to(device)
        prompt_embeds = [(k.to(device), v.repeat(batch_size, 1, 1).to(noise)) for k, v in prompt_embeds]
        negative_prompt_embeds = [(k.to(device), v.repeat(batch_size, 1, 1).to(noise)) for k, v in negative_prompt_embeds]
        pooled_prompt_embeds = pooled_prompt_embeds.repeat(batch_size, 1).to(noise)
        negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(batch_size, 1).to(noise)

        # Feeds

        sampler_kwargs = dict(
            cfg_scale=guidance_scale,
            positive=dict(
                encoder_hidden_states=prompt_embeds,
                added_cond_kwargs={"text_embeds": pooled_prompt_embeds, "time_ids": add_time_ids},
                cross_attention_kwargs=cross_attention_kwargs
            ),
            negative=dict(
                encoder_hidden_states=negative_prompt_embeds,
                added_cond_kwargs={"text_embeds": negative_pooled_prompt_embeds, "time_ids": add_neg_time_ids},
                cross_attention_kwargs=cross_attention_kwargs
            )
        )

        # Sample

        results = sample_dpmpp_2m(self.k_model, latents, sigmas, extra_args=sampler_kwargs, disable=False)

        return StableDiffusionXLPipelineOutput(images=results)