from typing import Callable, Union import math import torch from torch import Tensor import comfy.sample import comfy.model_patcher import comfy.utils from comfy.controlnet import ControlBase from comfy.model_patcher import ModelPatcher from comfy.ldm.modules.attention import BasicTransformerBlock from comfy.ldm.modules.diffusionmodules import openaimodel from .logger import logger from .utils import (AdvancedControlBase, ControlWeights, TimestepKeyframeGroup, AbstractPreprocWrapper, deepcopy_with_sharing, prepare_mask_batch, broadcast_image_to_extend) def refcn_sample_factory(orig_comfy_sample: Callable, is_custom=False) -> Callable: def get_refcn(control: ControlBase, order: int=-1): ref_set: set[ReferenceAdvanced] = set() if control is None: return ref_set if type(control) == ReferenceAdvanced: control.order = order order -= 1 ref_set.add(control) ref_set.update(get_refcn(control.previous_controlnet, order=order)) return ref_set def refcn_sample(model: ModelPatcher, *args, **kwargs): # check if positive or negative conds contain ref cn positive = args[-3] negative = args[-2] ref_set = set() if positive is not None: for cond in positive: if "control" in cond[1]: ref_set.update(get_refcn(cond[1]["control"])) if negative is not None: for cond in negative: if "control" in cond[1]: ref_set.update(get_refcn(cond[1]["control"])) # if no ref cn found, do original function immediately if len(ref_set) == 0: return orig_comfy_sample(model, *args, **kwargs) # otherwise, injection time try: # inject # storage for all Reference-related injections reference_injections = ReferenceInjections() # first, handle attn module injection all_modules = torch_dfs(model.model) attn_modules: list[RefBasicTransformerBlock] = [] for module in all_modules: if isinstance(module, BasicTransformerBlock): attn_modules.append(module) attn_modules = [module for module in all_modules if isinstance(module, BasicTransformerBlock)] attn_modules = sorted(attn_modules, key=lambda x: -x.norm1.normalized_shape[0]) for i, module in enumerate(attn_modules): injection_holder = InjectionBasicTransformerBlockHolder(block=module, idx=i) injection_holder.attn_weight = float(i) / float(len(attn_modules)) if hasattr(module, "_forward"): # backward compatibility module._forward = _forward_inject_BasicTransformerBlock.__get__(module, type(module)) else: module.forward = _forward_inject_BasicTransformerBlock.__get__(module, type(module)) module.injection_holder = injection_holder reference_injections.attn_modules.append(module) # figure out which module is middle block if hasattr(model.model.diffusion_model, "middle_block"): mid_modules = torch_dfs(model.model.diffusion_model.middle_block) mid_attn_modules: list[RefBasicTransformerBlock] = [module for module in mid_modules if isinstance(module, BasicTransformerBlock)] for module in mid_attn_modules: module.injection_holder.is_middle = True # next, handle gn module injection (TimestepEmbedSequential) # TODO: figure out the logic behind these hardcoded indexes if type(model.model).__name__ == "SDXL": input_block_indices = [4, 5, 7, 8] output_block_indices = [0, 1, 2, 3, 4, 5] else: input_block_indices = [4, 5, 7, 8, 10, 11] output_block_indices = [0, 1, 2, 3, 4, 5, 6, 7] if hasattr(model.model.diffusion_model, "middle_block"): module = model.model.diffusion_model.middle_block injection_holder = InjectionTimestepEmbedSequentialHolder(block=module, idx=0, is_middle=True) injection_holder.gn_weight = 0.0 module.injection_holder = injection_holder reference_injections.gn_modules.append(module) for w, i in enumerate(input_block_indices): module = model.model.diffusion_model.input_blocks[i] injection_holder = InjectionTimestepEmbedSequentialHolder(block=module, idx=i, is_input=True) injection_holder.gn_weight = 1.0 - float(w) / float(len(input_block_indices)) module.injection_holder = injection_holder reference_injections.gn_modules.append(module) for w, i in enumerate(output_block_indices): module = model.model.diffusion_model.output_blocks[i] injection_holder = InjectionTimestepEmbedSequentialHolder(block=module, idx=i, is_output=True) injection_holder.gn_weight = float(w) / float(len(output_block_indices)) module.injection_holder = injection_holder reference_injections.gn_modules.append(module) # hack gn_module forwards and update weights for i, module in enumerate(reference_injections.gn_modules): module.injection_holder.gn_weight *= 2 # handle diffusion_model forward injection reference_injections.diffusion_model_orig_forward = model.model.diffusion_model.forward model.model.diffusion_model.forward = factory_forward_inject_UNetModel(reference_injections).__get__(model.model.diffusion_model, type(model.model.diffusion_model)) # store ordered ref cns in model's transformer options orig_model_options = model.model_options new_model_options = model.model_options.copy() new_model_options["transformer_options"] = model.model_options["transformer_options"].copy() ref_list: list[ReferenceAdvanced] = list(ref_set) new_model_options["transformer_options"][REF_CONTROL_LIST_ALL] = sorted(ref_list, key=lambda x: x.order) model.model_options = new_model_options # continue with original function return orig_comfy_sample(model, *args, **kwargs) finally: # cleanup injections # restore attn modules attn_modules: list[RefBasicTransformerBlock] = reference_injections.attn_modules for module in attn_modules: module.injection_holder.restore(module) module.injection_holder.clean() del module.injection_holder del attn_modules # restore gn modules gn_modules: list[RefTimestepEmbedSequential] = reference_injections.gn_modules for module in gn_modules: module.injection_holder.restore(module) module.injection_holder.clean() del module.injection_holder del gn_modules # restore diffusion_model forward function model.model.diffusion_model.forward = reference_injections.diffusion_model_orig_forward.__get__(model.model.diffusion_model, type(model.model.diffusion_model)) # restore model_options model.model_options = orig_model_options # cleanup reference_injections.cleanup() return refcn_sample # inject sample functions comfy.sample.sample = refcn_sample_factory(comfy.sample.sample) comfy.sample.sample_custom = refcn_sample_factory(comfy.sample.sample_custom, is_custom=True) REF_ATTN_CONTROL_LIST = "ref_attn_control_list" REF_ADAIN_CONTROL_LIST = "ref_adain_control_list" REF_CONTROL_LIST_ALL = "ref_control_list_all" REF_CONTROL_INFO = "ref_control_info" REF_ATTN_MACHINE_STATE = "ref_attn_machine_state" REF_ADAIN_MACHINE_STATE = "ref_adain_machine_state" REF_COND_IDXS = "ref_cond_idxs" REF_UNCOND_IDXS = "ref_uncond_idxs" class MachineState: WRITE = "write" READ = "read" STYLEALIGN = "stylealign" OFF = "off" class ReferenceType: ATTN = "reference_attn" ADAIN = "reference_adain" ATTN_ADAIN = "reference_attn+adain" STYLE_ALIGN = "StyleAlign" _LIST = [ATTN, ADAIN, ATTN_ADAIN] _LIST_ATTN = [ATTN, ATTN_ADAIN] _LIST_ADAIN = [ADAIN, ATTN_ADAIN] @classmethod def is_attn(cls, ref_type: str): return ref_type in cls._LIST_ATTN @classmethod def is_adain(cls, ref_type: str): return ref_type in cls._LIST_ADAIN class ReferenceOptions: def __init__(self, reference_type: str, attn_style_fidelity: float, adain_style_fidelity: float, attn_ref_weight: float, adain_ref_weight: float, attn_strength: float=1.0, adain_strength: float=1.0, ref_with_other_cns: bool=False): self.reference_type = reference_type # attn self.original_attn_style_fidelity = attn_style_fidelity self.attn_style_fidelity = attn_style_fidelity self.attn_ref_weight = attn_ref_weight self.attn_strength = attn_strength # adain self.original_adain_style_fidelity = adain_style_fidelity self.adain_style_fidelity = adain_style_fidelity self.adain_ref_weight = adain_ref_weight self.adain_strength = adain_strength # other self.ref_with_other_cns = ref_with_other_cns def clone(self): return ReferenceOptions(reference_type=self.reference_type, attn_style_fidelity=self.original_attn_style_fidelity, adain_style_fidelity=self.original_adain_style_fidelity, attn_ref_weight=self.attn_ref_weight, adain_ref_weight=self.adain_ref_weight, attn_strength=self.attn_strength, adain_strength=self.adain_strength, ref_with_other_cns=self.ref_with_other_cns) @staticmethod def create_combo(reference_type: str, style_fidelity: float, ref_weight: float, ref_with_other_cns: bool=False): return ReferenceOptions(reference_type=reference_type, attn_style_fidelity=style_fidelity, adain_style_fidelity=style_fidelity, attn_ref_weight=ref_weight, adain_ref_weight=ref_weight, ref_with_other_cns=ref_with_other_cns) class ReferencePreprocWrapper(AbstractPreprocWrapper): error_msg = error_msg = "Invalid use of Reference Preprocess output. The output of RGB SparseCtrl preprocessor is NOT a usual image, but a latent pretending to be an image - you must connect the output directly to an Apply Advanced ControlNet node. It cannot be used for anything else that accepts IMAGE input." def __init__(self, condhint: Tensor): super().__init__(condhint) class ReferenceAdvanced(ControlBase, AdvancedControlBase): CHANNEL_TO_MULT = {320: 1, 640: 2, 1280: 4} def __init__(self, ref_opts: ReferenceOptions, timestep_keyframes: TimestepKeyframeGroup, device=None): super().__init__(device) AdvancedControlBase.__init__(self, super(), timestep_keyframes=timestep_keyframes, weights_default=ControlWeights.controllllite()) self.ref_opts = ref_opts self.order = 0 self.latent_format = None self.model_sampling_current = None self.should_apply_attn_effective_strength = False self.should_apply_adain_effective_strength = False self.should_apply_effective_masks = False self.latent_shape = None def any_attn_strength_to_apply(self): return self.should_apply_attn_effective_strength or self.should_apply_effective_masks def any_adain_strength_to_apply(self): return self.should_apply_adain_effective_strength or self.should_apply_effective_masks def get_effective_strength(self): effective_strength = self.strength if self._current_timestep_keyframe is not None: effective_strength = effective_strength * self._current_timestep_keyframe.strength return effective_strength def get_effective_attn_mask_or_float(self, x: Tensor, channels: int, is_mid: bool): if not self.should_apply_effective_masks: return self.get_effective_strength() * self.ref_opts.attn_strength if is_mid: div = 8 else: div = self.CHANNEL_TO_MULT[channels] real_mask = torch.ones([self.latent_shape[0], 1, self.latent_shape[2]//div, self.latent_shape[3]//div]).to(dtype=x.dtype, device=x.device) * self.strength * self.ref_opts.attn_strength self.apply_advanced_strengths_and_masks(x=real_mask, batched_number=self.batched_number) # mask is now shape [b, 1, h ,w]; need to turn into [b, h*w, 1] b, c, h, w = real_mask.shape real_mask = real_mask.permute(0, 2, 3, 1).reshape(b, h*w, c) return real_mask def get_effective_adain_mask_or_float(self, x: Tensor): if not self.should_apply_effective_masks: return self.get_effective_strength() * self.ref_opts.adain_strength b, c, h, w = x.shape real_mask = torch.ones([b, 1, h, w]).to(dtype=x.dtype, device=x.device) * self.strength * self.ref_opts.adain_strength self.apply_advanced_strengths_and_masks(x=real_mask, batched_number=self.batched_number) return real_mask def should_run(self): running = super().should_run() if not running: return running attn_run = False adain_run = False if ReferenceType.is_attn(self.ref_opts.reference_type): # attn will run as long as neither weight or strength is zero attn_run = not (math.isclose(self.ref_opts.attn_ref_weight, 0.0) or math.isclose(self.ref_opts.attn_strength, 0.0)) if ReferenceType.is_adain(self.ref_opts.reference_type): # adain will run as long as neither weight or strength is zero adain_run = not (math.isclose(self.ref_opts.adain_ref_weight, 0.0) or math.isclose(self.ref_opts.adain_strength, 0.0)) return attn_run or adain_run def pre_run_advanced(self, model, percent_to_timestep_function): AdvancedControlBase.pre_run_advanced(self, model, percent_to_timestep_function) if type(self.cond_hint_original) == ReferencePreprocWrapper: self.cond_hint_original = self.cond_hint_original.condhint self.latent_format = model.latent_format # LatentFormat object, used to process_in latent cond_hint self.model_sampling_current = model.model_sampling # SDXL is more sensitive to style_fidelity according to sd-webui-controlnet comments if type(model).__name__ == "SDXL": self.ref_opts.attn_style_fidelity = self.ref_opts.original_attn_style_fidelity ** 3.0 self.ref_opts.adain_style_fidelity = self.ref_opts.original_adain_style_fidelity ** 3.0 else: self.ref_opts.attn_style_fidelity = self.ref_opts.original_attn_style_fidelity self.ref_opts.adain_style_fidelity = self.ref_opts.original_adain_style_fidelity def get_control_advanced(self, x_noisy: Tensor, t, cond, batched_number: int): # normal ControlNet stuff control_prev = None if self.previous_controlnet is not None: control_prev = self.previous_controlnet.get_control(x_noisy, t, cond, batched_number) if self.timestep_range is not None: if t[0] > self.timestep_range[0] or t[0] < self.timestep_range[1]: return control_prev dtype = x_noisy.dtype # prepare cond_hint - it is a latent, NOT an image #if self.sub_idxs is not None or self.cond_hint is None or x_noisy.shape[2] != self.cond_hint.shape[2] or x_noisy.shape[3] != self.cond_hint.shape[3]: if self.cond_hint is not None: del self.cond_hint self.cond_hint = None # if self.cond_hint_original length greater or equal to real latent count, subdivide it before scaling if self.sub_idxs is not None and self.cond_hint_original.size(0) >= self.full_latent_length: self.cond_hint = comfy.utils.common_upscale( self.cond_hint_original[self.sub_idxs], x_noisy.shape[3], x_noisy.shape[2], 'nearest-exact', "center").to(dtype).to(self.device) else: self.cond_hint = comfy.utils.common_upscale( self.cond_hint_original, x_noisy.shape[3], x_noisy.shape[2], 'nearest-exact', "center").to(dtype).to(self.device) if x_noisy.shape[0] != self.cond_hint.shape[0]: self.cond_hint = broadcast_image_to_extend(self.cond_hint, x_noisy.shape[0], batched_number, except_one=False) # noise cond_hint based on sigma (current step) self.cond_hint = self.latent_format.process_in(self.cond_hint) self.cond_hint = ref_noise_latents(self.cond_hint, sigma=t, noise=None) timestep = self.model_sampling_current.timestep(t) self.should_apply_attn_effective_strength = not (math.isclose(self.strength, 1.0) and math.isclose(self._current_timestep_keyframe.strength, 1.0) and math.isclose(self.ref_opts.attn_strength, 1.0)) self.should_apply_adain_effective_strength = not (math.isclose(self.strength, 1.0) and math.isclose(self._current_timestep_keyframe.strength, 1.0) and math.isclose(self.ref_opts.adain_strength, 1.0)) # prepare mask - use direct_attn, so the mask dims will match source latents (and be smaller) self.prepare_mask_cond_hint(x_noisy=x_noisy, t=t, cond=cond, batched_number=batched_number, direct_attn=True) self.should_apply_effective_masks = self.latent_keyframes is not None or self.mask_cond_hint is not None or self.tk_mask_cond_hint is not None self.latent_shape = list(x_noisy.shape) # done preparing; model patches will take care of everything now. # return normal controlnet stuff return control_prev def cleanup_advanced(self): super().cleanup_advanced() del self.latent_format self.latent_format = None del self.model_sampling_current self.model_sampling_current = None self.should_apply_attn_effective_strength = False self.should_apply_adain_effective_strength = False self.should_apply_effective_masks = False def copy(self): c = ReferenceAdvanced(self.ref_opts, self.timestep_keyframes) c.order = self.order self.copy_to(c) self.copy_to_advanced(c) return c # avoid deepcopy shenanigans by making deepcopy not do anything to the reference # TODO: do the bookkeeping to do this in a proper way for all Adv-ControlNets def __deepcopy__(self, memo): return self def ref_noise_latents(latents: Tensor, sigma: Tensor, noise: Tensor=None): sigma = sigma.unsqueeze(-1).unsqueeze(-1).unsqueeze(-1) alpha_cumprod = 1 / ((sigma * sigma) + 1) sqrt_alpha_prod = alpha_cumprod ** 0.5 sqrt_one_minus_alpha_prod = (1. - alpha_cumprod) ** 0.5 if noise is None: # generator = torch.Generator(device="cuda") # generator.manual_seed(0) # noise = torch.empty_like(latents).normal_(generator=generator) # generator = torch.Generator() # generator.manual_seed(0) # noise = torch.randn(latents.size(), generator=generator).to(latents.device) noise = torch.randn_like(latents).to(latents.device) return sqrt_alpha_prod * latents + sqrt_one_minus_alpha_prod * noise def simple_noise_latents(latents: Tensor, sigma: float, noise: Tensor=None): if noise is None: noise = torch.rand_like(latents) return latents + noise * sigma class BankStylesBasicTransformerBlock: def __init__(self): self.bank = [] self.style_cfgs = [] self.cn_idx: list[int] = [] def get_avg_style_fidelity(self): return sum(self.style_cfgs) / float(len(self.style_cfgs)) def clean(self): del self.bank self.bank = [] del self.style_cfgs self.style_cfgs = [] del self.cn_idx self.cn_idx = [] class BankStylesTimestepEmbedSequential: def __init__(self): self.var_bank = [] self.mean_bank = [] self.style_cfgs = [] self.cn_idx: list[int] = [] def get_avg_var_bank(self): return sum(self.var_bank) / float(len(self.var_bank)) def get_avg_mean_bank(self): return sum(self.mean_bank) / float(len(self.mean_bank)) def get_avg_style_fidelity(self): return sum(self.style_cfgs) / float(len(self.style_cfgs)) def clean(self): del self.mean_bank self.mean_bank = [] del self.var_bank self.var_bank = [] del self.style_cfgs self.style_cfgs = [] del self.cn_idx self.cn_idx = [] class InjectionBasicTransformerBlockHolder: def __init__(self, block: BasicTransformerBlock, idx=None): if hasattr(block, "_forward"): # backward compatibility self.original_forward = block._forward else: self.original_forward = block.forward self.idx = idx self.attn_weight = 1.0 self.is_middle = False self.bank_styles = BankStylesBasicTransformerBlock() def restore(self, block: BasicTransformerBlock): if hasattr(block, "_forward"): # backward compatibility block._forward = self.original_forward else: block.forward = self.original_forward def clean(self): self.bank_styles.clean() class InjectionTimestepEmbedSequentialHolder: def __init__(self, block: openaimodel.TimestepEmbedSequential, idx=None, is_middle=False, is_input=False, is_output=False): self.original_forward = block.forward self.idx = idx self.gn_weight = 1.0 self.is_middle = is_middle self.is_input = is_input self.is_output = is_output self.bank_styles = BankStylesTimestepEmbedSequential() def restore(self, block: openaimodel.TimestepEmbedSequential): block.forward = self.original_forward def clean(self): self.bank_styles.clean() class ReferenceInjections: def __init__(self, attn_modules: list['RefBasicTransformerBlock']=None, gn_modules: list['RefTimestepEmbedSequential']=None): self.attn_modules = attn_modules if attn_modules else [] self.gn_modules = gn_modules if gn_modules else [] self.diffusion_model_orig_forward: Callable = None def clean_module_mem(self): for attn_module in self.attn_modules: try: attn_module.injection_holder.clean() except Exception: pass for gn_module in self.gn_modules: try: gn_module.injection_holder.clean() except Exception: pass def cleanup(self): self.clean_module_mem() del self.attn_modules self.attn_modules = [] del self.gn_modules self.gn_modules = [] self.diffusion_model_orig_forward = None def factory_forward_inject_UNetModel(reference_injections: ReferenceInjections): def forward_inject_UNetModel(self, x: Tensor, *args, **kwargs): # get control and transformer_options from kwargs real_args = list(args) real_kwargs = list(kwargs.keys()) control = kwargs.get("control", None) transformer_options = kwargs.get("transformer_options", None) # look for ReferenceAttnPatch objects to get ReferenceAdvanced objects ref_controlnets: list[ReferenceAdvanced] = transformer_options[REF_CONTROL_LIST_ALL] # discard any controlnets that should not run ref_controlnets = [x for x in ref_controlnets if x.should_run()] # if nothing related to reference controlnets, do nothing special if len(ref_controlnets) == 0: return reference_injections.diffusion_model_orig_forward(x, *args, **kwargs) try: # assign cond and uncond idxs batched_number = len(transformer_options["cond_or_uncond"]) per_batch = x.shape[0] // batched_number indiv_conds = [] for cond_type in transformer_options["cond_or_uncond"]: indiv_conds.extend([cond_type] * per_batch) transformer_options[REF_UNCOND_IDXS] = [i for i, x in enumerate(indiv_conds) if x == 1] transformer_options[REF_COND_IDXS] = [i for i, x in enumerate(indiv_conds) if x == 0] # check which controlnets do which thing attn_controlnets = [] adain_controlnets = [] for control in ref_controlnets: if ReferenceType.is_attn(control.ref_opts.reference_type): attn_controlnets.append(control) if ReferenceType.is_adain(control.ref_opts.reference_type): adain_controlnets.append(control) if len(adain_controlnets) > 0: # ComfyUI uses forward_timestep_embed with the TimestepEmbedSequential passed into it orig_forward_timestep_embed = openaimodel.forward_timestep_embed openaimodel.forward_timestep_embed = forward_timestep_embed_ref_inject_factory(orig_forward_timestep_embed) # handle running diffusion with ref cond hints for control in ref_controlnets: if ReferenceType.is_attn(control.ref_opts.reference_type): transformer_options[REF_ATTN_MACHINE_STATE] = MachineState.WRITE else: transformer_options[REF_ATTN_MACHINE_STATE] = MachineState.OFF if ReferenceType.is_adain(control.ref_opts.reference_type): transformer_options[REF_ADAIN_MACHINE_STATE] = MachineState.WRITE else: transformer_options[REF_ADAIN_MACHINE_STATE] = MachineState.OFF transformer_options[REF_ATTN_CONTROL_LIST] = [control] transformer_options[REF_ADAIN_CONTROL_LIST] = [control] orig_kwargs = kwargs if not control.ref_opts.ref_with_other_cns: kwargs = kwargs.copy() kwargs["control"] = None reference_injections.diffusion_model_orig_forward(control.cond_hint.to(dtype=x.dtype).to(device=x.device), *args, **kwargs) kwargs = orig_kwargs # run diffusion for real now transformer_options[REF_ATTN_MACHINE_STATE] = MachineState.READ transformer_options[REF_ADAIN_MACHINE_STATE] = MachineState.READ transformer_options[REF_ATTN_CONTROL_LIST] = attn_controlnets transformer_options[REF_ADAIN_CONTROL_LIST] = adain_controlnets return reference_injections.diffusion_model_orig_forward(x, *args, **kwargs) finally: # make sure banks are cleared no matter what happens - otherwise, RIP VRAM reference_injections.clean_module_mem() if len(adain_controlnets) > 0: openaimodel.forward_timestep_embed = orig_forward_timestep_embed return forward_inject_UNetModel # dummy class just to help IDE keep track of injected variables class RefBasicTransformerBlock(BasicTransformerBlock): injection_holder: InjectionBasicTransformerBlockHolder = None def _forward_inject_BasicTransformerBlock(self: RefBasicTransformerBlock, x: Tensor, context: Tensor=None, transformer_options: dict[str]={}): extra_options = {} block = transformer_options.get("block", None) block_index = transformer_options.get("block_index", 0) transformer_patches = {} transformer_patches_replace = {} for k in transformer_options: if k == "patches": transformer_patches = transformer_options[k] elif k == "patches_replace": transformer_patches_replace = transformer_options[k] else: extra_options[k] = transformer_options[k] extra_options["n_heads"] = self.n_heads extra_options["dim_head"] = self.d_head if self.ff_in: x_skip = x x = self.ff_in(self.norm_in(x)) if self.is_res: x += x_skip n: Tensor = self.norm1(x) if self.disable_self_attn: context_attn1 = context else: context_attn1 = None value_attn1 = None # Reference CN stuff uc_idx_mask = transformer_options.get(REF_UNCOND_IDXS, []) c_idx_mask = transformer_options.get(REF_COND_IDXS, []) # WRITE mode will only have one ReferenceAdvanced, other modes will have all ReferenceAdvanced ref_controlnets: list[ReferenceAdvanced] = transformer_options.get(REF_ATTN_CONTROL_LIST, None) ref_machine_state: str = transformer_options.get(REF_ATTN_MACHINE_STATE, None) # if in WRITE mode, save n and style_fidelity if ref_controlnets and ref_machine_state == MachineState.WRITE: if ref_controlnets[0].ref_opts.attn_ref_weight > self.injection_holder.attn_weight: self.injection_holder.bank_styles.bank.append(n.detach().clone()) self.injection_holder.bank_styles.style_cfgs.append(ref_controlnets[0].ref_opts.attn_style_fidelity) self.injection_holder.bank_styles.cn_idx.append(ref_controlnets[0].order) if "attn1_patch" in transformer_patches: patch = transformer_patches["attn1_patch"] if context_attn1 is None: context_attn1 = n value_attn1 = context_attn1 for p in patch: n, context_attn1, value_attn1 = p(n, context_attn1, value_attn1, extra_options) if block is not None: transformer_block = (block[0], block[1], block_index) else: transformer_block = None attn1_replace_patch = transformer_patches_replace.get("attn1", {}) block_attn1 = transformer_block if block_attn1 not in attn1_replace_patch: block_attn1 = block if block_attn1 in attn1_replace_patch: if context_attn1 is None: context_attn1 = n value_attn1 = n n = self.attn1.to_q(n) # Reference CN READ - use attn1_replace_patch appropriately if ref_machine_state == MachineState.READ and len(self.injection_holder.bank_styles.bank) > 0: bank_styles = self.injection_holder.bank_styles style_fidelity = bank_styles.get_avg_style_fidelity() real_bank = bank_styles.bank.copy() cn_idx = 0 for idx, order in enumerate(bank_styles.cn_idx): # make sure matching ref cn is selected for i in range(cn_idx, len(ref_controlnets)): if ref_controlnets[i].order == order: cn_idx = i break assert order == ref_controlnets[cn_idx].order if ref_controlnets[cn_idx].any_attn_strength_to_apply(): effective_strength = ref_controlnets[cn_idx].get_effective_attn_mask_or_float(x=n, channels=n.shape[2], is_mid=self.injection_holder.is_middle) real_bank[idx] = real_bank[idx] * effective_strength + context_attn1 * (1-effective_strength) n_uc = self.attn1.to_out(attn1_replace_patch[block_attn1]( n, self.attn1.to_k(torch.cat([context_attn1] + real_bank, dim=1)), self.attn1.to_v(torch.cat([value_attn1] + real_bank, dim=1)), extra_options)) n_c = n_uc.clone() if len(uc_idx_mask) > 0 and not math.isclose(style_fidelity, 0.0): n_c[uc_idx_mask] = self.attn1.to_out(attn1_replace_patch[block_attn1]( n[uc_idx_mask], self.attn1.to_k(context_attn1[uc_idx_mask]), self.attn1.to_v(value_attn1[uc_idx_mask]), extra_options)) n = style_fidelity * n_c + (1.0-style_fidelity) * n_uc bank_styles.clean() else: context_attn1 = self.attn1.to_k(context_attn1) value_attn1 = self.attn1.to_v(value_attn1) n = attn1_replace_patch[block_attn1](n, context_attn1, value_attn1, extra_options) n = self.attn1.to_out(n) else: # Reference CN READ - no attn1_replace_patch if ref_machine_state == MachineState.READ and len(self.injection_holder.bank_styles.bank) > 0: if context_attn1 is None: context_attn1 = n bank_styles = self.injection_holder.bank_styles style_fidelity = bank_styles.get_avg_style_fidelity() real_bank = bank_styles.bank.copy() cn_idx = 0 for idx, order in enumerate(bank_styles.cn_idx): # make sure matching ref cn is selected for i in range(cn_idx, len(ref_controlnets)): if ref_controlnets[i].order == order: cn_idx = i break assert order == ref_controlnets[cn_idx].order if ref_controlnets[cn_idx].any_attn_strength_to_apply(): effective_strength = ref_controlnets[cn_idx].get_effective_attn_mask_or_float(x=n, channels=n.shape[2], is_mid=self.injection_holder.is_middle) real_bank[idx] = real_bank[idx] * effective_strength + context_attn1 * (1-effective_strength) n_uc: Tensor = self.attn1( n, context=torch.cat([context_attn1] + real_bank, dim=1), value=torch.cat([value_attn1] + real_bank, dim=1) if value_attn1 is not None else value_attn1) n_c = n_uc.clone() if len(uc_idx_mask) > 0 and not math.isclose(style_fidelity, 0.0): n_c[uc_idx_mask] = self.attn1( n[uc_idx_mask], context=context_attn1[uc_idx_mask], value=value_attn1[uc_idx_mask] if value_attn1 is not None else value_attn1) n = style_fidelity * n_c + (1.0-style_fidelity) * n_uc bank_styles.clean() else: n = self.attn1(n, context=context_attn1, value=value_attn1) if "attn1_output_patch" in transformer_patches: patch = transformer_patches["attn1_output_patch"] for p in patch: n = p(n, extra_options) x += n if "middle_patch" in transformer_patches: patch = transformer_patches["middle_patch"] for p in patch: x = p(x, extra_options) if self.attn2 is not None: n = self.norm2(x) if self.switch_temporal_ca_to_sa: context_attn2 = n else: context_attn2 = context value_attn2 = None if "attn2_patch" in transformer_patches: patch = transformer_patches["attn2_patch"] value_attn2 = context_attn2 for p in patch: n, context_attn2, value_attn2 = p(n, context_attn2, value_attn2, extra_options) attn2_replace_patch = transformer_patches_replace.get("attn2", {}) block_attn2 = transformer_block if block_attn2 not in attn2_replace_patch: block_attn2 = block if block_attn2 in attn2_replace_patch: if value_attn2 is None: value_attn2 = context_attn2 n = self.attn2.to_q(n) context_attn2 = self.attn2.to_k(context_attn2) value_attn2 = self.attn2.to_v(value_attn2) n = attn2_replace_patch[block_attn2](n, context_attn2, value_attn2, extra_options) n = self.attn2.to_out(n) else: n = self.attn2(n, context=context_attn2, value=value_attn2) if "attn2_output_patch" in transformer_patches: patch = transformer_patches["attn2_output_patch"] for p in patch: n = p(n, extra_options) x += n if self.is_res: x_skip = x x = self.ff(self.norm3(x)) if self.is_res: x += x_skip return x class RefTimestepEmbedSequential(openaimodel.TimestepEmbedSequential): injection_holder: InjectionTimestepEmbedSequentialHolder = None def forward_timestep_embed_ref_inject_factory(orig_timestep_embed_inject_factory: Callable): def forward_timestep_embed_ref_inject(*args, **kwargs): ts: RefTimestepEmbedSequential = args[0] if not hasattr(ts, "injection_holder"): return orig_timestep_embed_inject_factory(*args, **kwargs) eps = 1e-6 x: Tensor = orig_timestep_embed_inject_factory(*args, **kwargs) y: Tensor = None transformer_options: dict[str] = args[4] # Reference CN stuff uc_idx_mask = transformer_options.get(REF_UNCOND_IDXS, []) c_idx_mask = transformer_options.get(REF_COND_IDXS, []) # WRITE mode will only have one ReferenceAdvanced, other modes will have all ReferenceAdvanced ref_controlnets: list[ReferenceAdvanced] = transformer_options.get(REF_ADAIN_CONTROL_LIST, None) ref_machine_state: str = transformer_options.get(REF_ADAIN_MACHINE_STATE, None) # if in WRITE mode, save var, mean, and style_cfg if ref_machine_state == MachineState.WRITE: if ref_controlnets[0].ref_opts.adain_ref_weight > ts.injection_holder.gn_weight: var, mean = torch.var_mean(x, dim=(2, 3), keepdim=True, correction=0) ts.injection_holder.bank_styles.var_bank.append(var) ts.injection_holder.bank_styles.mean_bank.append(mean) ts.injection_holder.bank_styles.style_cfgs.append(ref_controlnets[0].ref_opts.adain_style_fidelity) ts.injection_holder.bank_styles.cn_idx.append(ref_controlnets[0].order) # if in READ mode, do math with saved var, mean, and style_cfg if ref_machine_state == MachineState.READ: if len(ts.injection_holder.bank_styles.var_bank) > 0: bank_styles = ts.injection_holder.bank_styles var, mean = torch.var_mean(x, dim=(2, 3), keepdim=True, correction=0) std = torch.maximum(var, torch.zeros_like(var) + eps) ** 0.5 y_uc = torch.zeros_like(x) cn_idx = 0 for idx, order in enumerate(bank_styles.cn_idx): # make sure matching ref cn is selected for i in range(cn_idx, len(ref_controlnets)): if ref_controlnets[i].order == order: cn_idx = i break assert order == ref_controlnets[cn_idx].order style_fidelity = bank_styles.style_cfgs[idx] var_acc = bank_styles.var_bank[idx] mean_acc = bank_styles.mean_bank[idx] std_acc = torch.maximum(var_acc, torch.zeros_like(var_acc) + eps) ** 0.5 sub_y_uc = (((x - mean) / std) * std_acc) + mean_acc if ref_controlnets[cn_idx].any_adain_strength_to_apply(): effective_strength = ref_controlnets[cn_idx].get_effective_adain_mask_or_float(x=x) sub_y_uc = sub_y_uc * effective_strength + x * (1-effective_strength) y_uc += sub_y_uc # get average, if more than one if len(bank_styles.cn_idx) > 1: y_uc /= len(bank_styles.cn_idx) y_c = y_uc.clone() if len(uc_idx_mask) > 0 and not math.isclose(style_fidelity, 0.0): y_c[uc_idx_mask] = x.to(y_c.dtype)[uc_idx_mask] y = style_fidelity * y_c + (1.0 - style_fidelity) * y_uc ts.injection_holder.bank_styles.clean() if y is None: y = x return y.to(x.dtype) return forward_timestep_embed_ref_inject # DFS Search for Torch.nn.Module, Written by Lvmin def torch_dfs(model: torch.nn.Module): result = [model] for child in model.children(): result += torch_dfs(child) return result