diffsynth 1.0.0__py3-none-any.whl

diffsynth/prompters/sdxl_prompter.py

@@ -0,0 +1,61 @@
+from .base_prompter import BasePrompter, tokenize_long_prompt
+from ..models.model_manager import ModelManager
+from ..models import SDXLTextEncoder, SDXLTextEncoder2
+from transformers import CLIPTokenizer
+import torch, os
+
+
+
+class SDXLPrompter(BasePrompter):
+    def __init__(
+        self,
+        tokenizer_path=None,
+        tokenizer_2_path=None
+    ):
+        if tokenizer_path is None:
+            base_path = os.path.dirname(os.path.dirname(__file__))
+            tokenizer_path = os.path.join(base_path, "tokenizer_configs/stable_diffusion/tokenizer")
+        if tokenizer_2_path is None:
+            base_path = os.path.dirname(os.path.dirname(__file__))
+            tokenizer_2_path = os.path.join(base_path, "tokenizer_configs/stable_diffusion_xl/tokenizer_2")
+        super().__init__()
+        self.tokenizer = CLIPTokenizer.from_pretrained(tokenizer_path)
+        self.tokenizer_2 = CLIPTokenizer.from_pretrained(tokenizer_2_path)
+        self.text_encoder: SDXLTextEncoder = None
+        self.text_encoder_2: SDXLTextEncoder2 = None
+
+    
+    def fetch_models(self, text_encoder: SDXLTextEncoder = None, text_encoder_2: SDXLTextEncoder2 = None):
+        self.text_encoder = text_encoder
+        self.text_encoder_2 = text_encoder_2
+    
+    
+    def encode_prompt(
+        self,
+        prompt,
+        clip_skip=1,
+        clip_skip_2=2,
+        positive=True,
+        device="cuda"
+    ):
+        prompt = self.process_prompt(prompt, positive=positive)
+        
+        # 1
+        input_ids = tokenize_long_prompt(self.tokenizer, prompt).to(device)
+        prompt_emb_1 = self.text_encoder(input_ids, clip_skip=clip_skip)
+
+        # 2
+        input_ids_2 = tokenize_long_prompt(self.tokenizer_2, prompt).to(device)
+        add_text_embeds, prompt_emb_2 = self.text_encoder_2(input_ids_2, clip_skip=clip_skip_2)
+
+        # Merge
+        if prompt_emb_1.shape[0] != prompt_emb_2.shape[0]:
+            max_batch_size = min(prompt_emb_1.shape[0], prompt_emb_2.shape[0])
+            prompt_emb_1 = prompt_emb_1[: max_batch_size]
+            prompt_emb_2 = prompt_emb_2[: max_batch_size]
+        prompt_emb = torch.concatenate([prompt_emb_1, prompt_emb_2], dim=-1)
+
+        # For very long prompt, we only use the first 77 tokens to compute `add_text_embeds`.
+        add_text_embeds = add_text_embeds[0:1]
+        prompt_emb = prompt_emb.reshape((1, prompt_emb.shape[0]*prompt_emb.shape[1], -1))
+        return add_text_embeds, prompt_emb

diffsynth/schedulers/__init__.py

@@ -0,0 +1,3 @@
+from .ddim import EnhancedDDIMScheduler
+from .continuous_ode import ContinuousODEScheduler
+from .flow_match import FlowMatchScheduler

diffsynth/schedulers/continuous_ode.py

@@ -0,0 +1,59 @@
+import torch
+
+
+class ContinuousODEScheduler():
+
+    def __init__(self, num_inference_steps=100, sigma_max=700.0, sigma_min=0.002, rho=7.0):
+        self.sigma_max = sigma_max
+        self.sigma_min = sigma_min
+        self.rho = rho
+        self.set_timesteps(num_inference_steps)
+
+
+    def set_timesteps(self, num_inference_steps=100, denoising_strength=1.0):
+        ramp = torch.linspace(1-denoising_strength, 1, num_inference_steps)
+        min_inv_rho = torch.pow(torch.tensor((self.sigma_min,)), (1 / self.rho))
+        max_inv_rho = torch.pow(torch.tensor((self.sigma_max,)), (1 / self.rho))
+        self.sigmas = torch.pow(max_inv_rho + ramp * (min_inv_rho - max_inv_rho), self.rho)
+        self.timesteps = torch.log(self.sigmas) * 0.25
+
+
+    def step(self, model_output, timestep, sample, to_final=False):
+        timestep_id = torch.argmin((self.timesteps - timestep).abs())
+        sigma = self.sigmas[timestep_id]
+        sample *= (sigma*sigma + 1).sqrt()
+        estimated_sample = -sigma / (sigma*sigma + 1).sqrt() * model_output + 1 / (sigma*sigma + 1) * sample
+        if to_final or timestep_id + 1 >= len(self.timesteps):
+            prev_sample = estimated_sample
+        else:
+            sigma_ = self.sigmas[timestep_id + 1]
+            derivative = 1 / sigma * (sample - estimated_sample)
+            prev_sample = sample + derivative * (sigma_ - sigma)
+            prev_sample /= (sigma_*sigma_ + 1).sqrt()
+        return prev_sample
+    
+
+    def return_to_timestep(self, timestep, sample, sample_stablized):
+        # This scheduler doesn't support this function.
+        pass
+    
+    
+    def add_noise(self, original_samples, noise, timestep):
+        timestep_id = torch.argmin((self.timesteps - timestep).abs())
+        sigma = self.sigmas[timestep_id]
+        sample = (original_samples + noise * sigma) / (sigma*sigma + 1).sqrt()
+        return sample
+    
+
+    def training_target(self, sample, noise, timestep):
+        timestep_id = torch.argmin((self.timesteps - timestep).abs())
+        sigma = self.sigmas[timestep_id]
+        target = (-(sigma*sigma + 1).sqrt() / sigma + 1 / (sigma*sigma + 1).sqrt() / sigma) * sample + 1 / (sigma*sigma + 1).sqrt() * noise
+        return target
+    
+
+    def training_weight(self, timestep):
+        timestep_id = torch.argmin((self.timesteps - timestep).abs())
+        sigma = self.sigmas[timestep_id]
+        weight = (1 + sigma*sigma).sqrt() / sigma
+        return weight

diffsynth/schedulers/ddim.py

@@ -0,0 +1,79 @@
+import torch, math
+
+
+class EnhancedDDIMScheduler():
+
+    def __init__(self, num_train_timesteps=1000, beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear", prediction_type="epsilon"):
+        self.num_train_timesteps = num_train_timesteps
+        if beta_schedule == "scaled_linear":
+            betas = torch.square(torch.linspace(math.sqrt(beta_start), math.sqrt(beta_end), num_train_timesteps, dtype=torch.float32))
+        elif beta_schedule == "linear":
+            betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32)
+        else:
+            raise NotImplementedError(f"{beta_schedule} is not implemented")
+        self.alphas_cumprod = torch.cumprod(1.0 - betas, dim=0).tolist()
+        self.set_timesteps(10)
+        self.prediction_type = prediction_type
+
+
+    def set_timesteps(self, num_inference_steps, denoising_strength=1.0):
+        # The timesteps are aligned to 999...0, which is different from other implementations,
+        # but I think this implementation is more reasonable in theory.
+        max_timestep = max(round(self.num_train_timesteps * denoising_strength) - 1, 0)
+        num_inference_steps = min(num_inference_steps, max_timestep + 1)
+        if num_inference_steps == 1:
+            self.timesteps = torch.Tensor([max_timestep])
+        else:
+            step_length = max_timestep / (num_inference_steps - 1)
+            self.timesteps = torch.Tensor([round(max_timestep - i*step_length) for i in range(num_inference_steps)])
+
+
+    def denoise(self, model_output, sample, alpha_prod_t, alpha_prod_t_prev):
+        if self.prediction_type == "epsilon":
+            weight_e = math.sqrt(1 - alpha_prod_t_prev) - math.sqrt(alpha_prod_t_prev * (1 - alpha_prod_t) / alpha_prod_t)
+            weight_x = math.sqrt(alpha_prod_t_prev / alpha_prod_t)
+            prev_sample = sample * weight_x + model_output * weight_e
+        elif self.prediction_type == "v_prediction":
+            weight_e = -math.sqrt(alpha_prod_t_prev * (1 - alpha_prod_t)) + math.sqrt(alpha_prod_t * (1 - alpha_prod_t_prev))
+            weight_x = math.sqrt(alpha_prod_t * alpha_prod_t_prev) + math.sqrt((1 - alpha_prod_t) * (1 - alpha_prod_t_prev))
+            prev_sample = sample * weight_x + model_output * weight_e
+        else:
+            raise NotImplementedError(f"{self.prediction_type} is not implemented")
+        return prev_sample
+
+
+    def step(self, model_output, timestep, sample, to_final=False):
+        alpha_prod_t = self.alphas_cumprod[int(timestep.flatten().tolist()[0])]
+        if isinstance(timestep, torch.Tensor):
+            timestep = timestep.cpu()
+        timestep_id = torch.argmin((self.timesteps - timestep).abs())
+        if to_final or timestep_id + 1 >= len(self.timesteps):
+            alpha_prod_t_prev = 1.0
+        else:
+            timestep_prev = int(self.timesteps[timestep_id + 1])
+            alpha_prod_t_prev = self.alphas_cumprod[timestep_prev]
+
+        return self.denoise(model_output, sample, alpha_prod_t, alpha_prod_t_prev)
+
+
+    def return_to_timestep(self, timestep, sample, sample_stablized):
+        alpha_prod_t = self.alphas_cumprod[int(timestep.flatten().tolist()[0])]
+        noise_pred = (sample - math.sqrt(alpha_prod_t) * sample_stablized) / math.sqrt(1 - alpha_prod_t)
+        return noise_pred
+    
+    
+    def add_noise(self, original_samples, noise, timestep):
+        sqrt_alpha_prod = math.sqrt(self.alphas_cumprod[int(timestep.flatten().tolist()[0])])
+        sqrt_one_minus_alpha_prod = math.sqrt(1 - self.alphas_cumprod[int(timestep.flatten().tolist()[0])])
+        noisy_samples = sqrt_alpha_prod * original_samples + sqrt_one_minus_alpha_prod * noise
+        return noisy_samples
+    
+
+    def training_target(self, sample, noise, timestep):
+        if self.prediction_type == "epsilon":
+            return noise
+        else:
+            sqrt_alpha_prod = math.sqrt(self.alphas_cumprod[int(timestep.flatten().tolist()[0])])
+            sqrt_one_minus_alpha_prod = math.sqrt(1 - self.alphas_cumprod[int(timestep.flatten().tolist()[0])])
+            target = sqrt_alpha_prod * noise - sqrt_one_minus_alpha_prod * sample
+            return target

diffsynth/schedulers/flow_match.py

@@ -0,0 +1,51 @@
+import torch
+
+
+
+class FlowMatchScheduler():
+
+    def __init__(self, num_inference_steps=100, num_train_timesteps=1000, shift=3.0, sigma_max=1.0, sigma_min=0.003/1.002):
+        self.num_train_timesteps = num_train_timesteps
+        self.shift = shift
+        self.sigma_max = sigma_max
+        self.sigma_min = sigma_min
+        self.set_timesteps(num_inference_steps)
+
+
+    def set_timesteps(self, num_inference_steps=100, denoising_strength=1.0):
+        sigma_start = self.sigma_min + (self.sigma_max - self.sigma_min) * denoising_strength
+        self.sigmas = torch.linspace(sigma_start, self.sigma_min, num_inference_steps)
+        self.sigmas = self.shift * self.sigmas / (1 + (self.shift - 1) * self.sigmas)
+        self.timesteps = self.sigmas * self.num_train_timesteps
+
+
+    def step(self, model_output, timestep, sample, to_final=False):
+        if isinstance(timestep, torch.Tensor):
+            timestep = timestep.cpu()
+        timestep_id = torch.argmin((self.timesteps - timestep).abs())
+        sigma = self.sigmas[timestep_id]
+        if to_final or timestep_id + 1 >= len(self.timesteps):
+            sigma_ = 0
+        else:
+            sigma_ = self.sigmas[timestep_id + 1]
+        prev_sample = sample + model_output * (sigma_ - sigma)
+        return prev_sample
+    
+
+    def return_to_timestep(self, timestep, sample, sample_stablized):
+        # This scheduler doesn't support this function.
+        pass
+    
+    
+    def add_noise(self, original_samples, noise, timestep):
+        if isinstance(timestep, torch.Tensor):
+            timestep = timestep.cpu()
+        timestep_id = torch.argmin((self.timesteps - timestep).abs())
+        sigma = self.sigmas[timestep_id]
+        sample = (1 - sigma) * original_samples + sigma * noise
+        return sample
+    
+
+    def training_target(self, sample, noise, timestep):
+        target = noise - sample
+        return target

diffsynth/tokenizer_configs/hunyuan_dit/tokenizer/special_tokens_map.json

@@ -0,0 +1,7 @@
+{
+  "cls_token": "[CLS]",
+  "mask_token": "[MASK]",
+  "pad_token": "[PAD]",
+  "sep_token": "[SEP]",
+  "unk_token": "[UNK]"
+}

diffsynth/tokenizer_configs/hunyuan_dit/tokenizer/tokenizer_config.json

@@ -0,0 +1,16 @@
+{
+  "cls_token": "[CLS]",
+  "do_basic_tokenize": true,
+  "do_lower_case": true,
+  "mask_token": "[MASK]",
+  "name_or_path": "hfl/chinese-roberta-wwm-ext",
+  "never_split": null,
+  "pad_token": "[PAD]",
+  "sep_token": "[SEP]",
+  "special_tokens_map_file": "/home/chenweifeng/.cache/huggingface/hub/models--hfl--chinese-roberta-wwm-ext/snapshots/5c58d0b8ec1d9014354d691c538661bf00bfdb44/special_tokens_map.json",
+  "strip_accents": null,
+  "tokenize_chinese_chars": true,
+  "tokenizer_class": "BertTokenizer",
+  "unk_token": "[UNK]",
+  "model_max_length": 77
+}