init

utils/module_utils.py

@@ -0,0 +1,73 @@
+import torch
+
+from ..modules.attn_wrapper import get_attn_wrapper
+from ..modules.ref_controller import RefController
+
+
+def isinstance_str(x: object, cls_name: str):
+    for _cls in x.__class__.__mro__:
+        if _cls.__name__ == cls_name:
+            return True
+    
+    return False
+
+
+def is_transformer_block(module):
+    return isinstance_str(module, 'BasicTransformerBlock')
+
+
+def is_temporal_block(module):
+    return isinstance_str(module, 'TemporalTransformerBlock')
+
+
+def is_named_module_transformer_block(named_module):
+    if is_transformer_block(named_module[1]):
+        return True
+    elif is_temporal_block(named_module[1]):
+        return True
+    return False
+
+
+def get_norm1_shape(block_module):
+    if is_transformer_block(block_module):
+        return block_module.norm1.normalized_shape[0]
+    elif is_temporal_block(block_module):
+        return block_module.norms[0].normalized_shape[0]
+    raise ValueError('Bad block!', block_module)
+
+
+def get_block_attention(block_module):
+    if is_transformer_block(block_module):
+        return block_module.attn1
+    elif is_temporal_block(block_module):
+        return block_module.attention_blocks[0]
+
+
+def sort_block(module):
+    return -get_norm1_shape(module)
+
+
+def setup_ref_attn(model, attn_type):
+    # TODO: attn_type for full/mid_out
+    model = model.model.diffusion_model
+    block_modules = list(filter(is_named_module_transformer_block, model.named_modules()))
+    block_modules = list(map(lambda x: x[1], block_modules))
+    block_modules = sorted(block_modules, key=sort_block)
+
+    ref_controller = RefController()
+
+    for i, block_module in enumerate(block_modules):
+        if is_transformer_block(block_module):
+            attn = get_block_attention(block_module)
+            if not hasattr(attn, 'ref_mode'):
+                attn.__class__ = get_attn_wrapper(attn.__class__)
+            attn.ref_attn_weight = float(i) / float(len(block_modules))
+            ref_controller.add_module(attn)
+
+
+    temp_transformers = list(filter(lambda x: isinstance_str(x, 'TemporalTransformer3DModel'), model.named_modules()))
+
+    for _, transformer_module in temp_transformers:
+        ref_controller.add_temporal_transformer(transformer_module)
+
+    return ref_controller

utils/noise_utils.py

@@ -0,0 +1,22 @@
+import torch 
+
+
+def get_alphacumprod(sigma):
+    return 1 / ((sigma * sigma) + 1)
+
+
+def add_noise(src_latent, noise, sigma):
+    alphas_cumprod = get_alphacumprod(sigma)
+
+    sqrt_alpha_prod = alphas_cumprod ** 0.5
+    sqrt_alpha_prod = sqrt_alpha_prod.flatten()
+    while len(sqrt_alpha_prod.shape) < len(src_latent.shape):
+        sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1)
+
+    sqrt_one_minus_alpha_prod = (1 - alphas_cumprod) ** 0.5
+    sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten()
+    while len(sqrt_one_minus_alpha_prod.shape) < len(src_latent.shape):
+        sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1)
+
+    noisy_samples = sqrt_alpha_prod * src_latent + sqrt_one_minus_alpha_prod * noise
+    return noisy_samples

utils/ref_config.py

@@ -0,0 +1,7 @@
+
+
+
+class RefConfig:
+    def __init__(self, style_fidelity, attention_bound):
+        self.style_fidelity = style_fidelity
+        self.attention_bound = attention_bound

utils/ref_sampling_utils.py

@@ -0,0 +1,93 @@
+import torch
+
+import comfy.model_management
+
+from .ref_config import RefConfig
+from ..modules.ref_controller import RefController, RefMode
+from ..utils.noise_utils import add_noise
+
+
+def get_base_model(model):
+    base_model = model
+    while hasattr(base_model, 'inner_model'):
+        base_model = base_model.inner_model
+    return base_model
+
+
+def prepare_ref_latents(model, ref_latent):
+    base_model = get_base_model(model)
+    ref_latent = ref_latent['samples']
+    ref_latent = base_model.process_latent_in(ref_latent)
+    device = comfy.model_management.get_torch_device()
+    ref_latent = ref_latent.to(device)
+    return ref_latent
+
+
+def create_model_wrapper(model, sigmas, ref_latent, extra_args, ref_config:RefConfig):
+    # extra_args: ['cond', 'uncond', 'cond_scale', 'model_options', 'seed', 'denoise_mask']
+    ref_latent = prepare_ref_latents(model, ref_latent)
+
+    # Get ref_controller
+    model_options = extra_args.get('model_options', {})
+    ref_controller: RefController = model_options.get('ref_controller', {})
+
+    ref_conds = []
+    for cond in extra_args['cond']:
+        cond = {**cond}
+        if 'control' in cond:
+            del cond['control']
+        ref_conds.append(cond)
+
+    ref_unconds = []
+    for cond in extra_args['uncond']:
+        cond = {**cond}
+        if 'control' in cond:
+            del cond['control']
+        ref_unconds.append(cond)
+    
+    # Setup ref
+    ref_controller.clear_modules()
+    is_cfg = extra_args.get('cond_scale', 7) > 1
+    ref_controller.set_cfg(is_cfg)
+    ref_controller.set_style_fidelity(ref_config.style_fidelity)
+    ref_controller.set_attention_bound(ref_config.attention_bound)
+
+    def model_sample(latents, sigma, **kwargs):
+        ref_controller.get_default_temporal_values()
+        # Do Ref Sampling
+        ref_controller.set_temporal_values_to(1, 1)
+        ref_controller.set_mode(RefMode.WRITE)
+        ref_latent_noised = add_noise(ref_latent, torch.randn_like(ref_latent), sigma[0])
+        ref_sigma = torch.tensor([sigma[0]]*len(ref_latent)).to(sigmas.device)
+
+        model(ref_latent_noised, ref_sigma, **{**kwargs, 'cond': ref_conds, 'uncond': ref_unconds})
+
+        # Do Latent Sampling
+        ref_controller.set_temporal_values_to_default()
+        ref_controller.set_mode(RefMode.READ)
+        latents = model(latents, sigma, **kwargs)
+
+        # Clear refs
+        ref_controller.clear_modules()
+        ref_controller.set_mode(RefMode.OFF)
+        return latents
+
+
+    class RefModelWrapper:
+        inner_model = model.inner_model
+        def __call__(self, *args, **kwargs):
+            return model_sample(*args, **kwargs)
+        
+    return RefModelWrapper()
+
+
+def create_sampler(sample_fn, inversion_latent, ref_config):
+    @torch.no_grad()
+    def sample(model, latents, sigmas, extra_args=None, callback=None, disable=None, **extra_options):
+        # src_latents = prepare_src_latents(model, src_latent_image)
+        model_wrapper = create_model_wrapper(model, sigmas, inversion_latent, extra_args, ref_config)
+        output = sample_fn(model_wrapper, latents, sigmas, extra_args=extra_args, callback=callback, disable=disable)
+        del model_wrapper
+        return output
+    
+    return sample