diff --git a/adv_control/control_reference.py b/adv_control/control_reference.py index 28f1105..c36d747 100644 --- a/adv_control/control_reference.py +++ b/adv_control/control_reference.py @@ -17,6 +17,11 @@ from .utils import (AdvancedControlBase, ControlWeights, TimestepKeyframeGroup, broadcast_image_to_extend) +REF_READ_ATTN_CONTROL_LIST = "ref_read_attn_control_list" +REF_WRITE_ATTN_CONTROL_LIST = "ref_write_attn_control_list" +REF_READ_ADAIN_CONTROL_LIST = "ref_read_adain_control_list" +REF_WRITE_ADAIN_CONTROL_LIST = "ref_write_adain_control_list" + REF_ATTN_CONTROL_LIST = "ref_attn_control_list" REF_ADAIN_CONTROL_LIST = "ref_adain_control_list" REF_CONTROL_LIST_ALL = "ref_control_list_all" @@ -26,6 +31,13 @@ REF_ADAIN_MACHINE_STATE = "ref_adain_machine_state" REF_COND_IDXS = "ref_cond_idxs" REF_UNCOND_IDXS = "ref_uncond_idxs" +CONTEXTREF_OPTIONS_CLASS = "contextref_options_class" +CONTEXTREF_CLEAN_FUNC = "contextref_clean_func" +CONTEXTREF_CONTROL_LIST_ALL = "contextref_control_list_all" +CONTEXTREF_MACHINE_STATE = "contextref_machine_state" +CONTEXTREF_ATTN_MACHINE_STATE = "contextref_attn_machine_state" +CONTEXTREF_ADAIN_MACHINE_STATE = "contextref_adain_machine_state" + class MachineState: WRITE = "write" @@ -111,6 +123,8 @@ class ReferenceAdvanced(ControlBase, AdvancedControlBase): self.should_apply_adain_effective_strength = False self.should_apply_effective_masks = False self.latent_shape = None + # ContextRef stuff + self.is_context_ref = False def any_attn_strength_to_apply(self): return self.should_apply_attn_effective_strength or self.should_apply_effective_masks @@ -185,25 +199,27 @@ class ReferenceAdvanced(ControlBase, AdvancedControlBase): 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.model_latent_format.process_in(self.cond_hint) - self.cond_hint = ref_noise_latents(self.cond_hint, sigma=t, noise=None) + # cond_hint_original only matters for RefCN, NOT ContextRef + if self.cond_hint_original is not None: + # 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.model_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)) @@ -228,6 +244,7 @@ class ReferenceAdvanced(ControlBase, AdvancedControlBase): def copy(self): c = ReferenceAdvanced(self.ref_opts, self.timestep_keyframes) c.order = self.order + c.is_context_ref = self.is_context_ref self.copy_to(c) self.copy_to_advanced(c) return c @@ -238,6 +255,18 @@ class ReferenceAdvanced(ControlBase, AdvancedControlBase): return self +def handle_context_ref_setup(transformer_options): + transformer_options[CONTEXTREF_ATTN_MACHINE_STATE] = MachineState.OFF + transformer_options[CONTEXTREF_ADAIN_MACHINE_STATE] = MachineState.OFF + opts = ReferenceOptions(ReferenceType.ATTN, attn_style_fidelity=1.0, attn_ref_weight=1.0, attn_strength=1.0, adain_style_fidelity=0.0, adain_ref_weight=0.0) + cref = ReferenceAdvanced(ref_opts=opts, timestep_keyframes=None) + cref.order = -1 + context_ref_list = [cref] + transformer_options[CONTEXTREF_CONTROL_LIST_ALL] = context_ref_list + transformer_options[CONTEXTREF_OPTIONS_CLASS] = ReferenceOptions + return context_ref_list + + 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) @@ -262,48 +291,109 @@ def simple_noise_latents(latents: Tensor, sigma: float, noise: Tensor=None): class BankStylesBasicTransformerBlock: def __init__(self): + # ref self.bank = [] self.style_cfgs = [] self.cn_idx: list[int] = [] + # contextref + self.c_bank = [] + self.c_style_cfgs = [] + self.c_cn_idx = [] + + def get_bank(self, ignore_contextref=False): + if ignore_contextref: + return self.bank + return self.bank + self.c_bank + + def get_avg_style_fidelity(self, ignore_contextref=False): + if ignore_contextref: + return sum(self.style_cfgs) / float(len(self.style_cfgs)) + combined = self.style_cfgs + self.c_style_cfgs + return sum(combined) / float(len(combined)) - def get_avg_style_fidelity(self): - return sum(self.style_cfgs) / float(len(self.style_cfgs)) - - def clean(self): + def get_cn_idxs(self, ignore_contxtref=False): + if ignore_contxtref: + return self.cn_idx + return self.cn_idx + self.c_cn_idx + + def clean_ref(self): del self.bank - self.bank = [] del self.style_cfgs - self.style_cfgs = [] del self.cn_idx + self.bank = [] + self.style_cfgs = [] self.cn_idx = [] + def clean_contextref(self): + del self.c_bank + del self.c_style_cfgs + del self.c_cn_idx + self.c_bank = [] + self.c_style_cfgs = [] + self.c_cn_idx = [] + + def clean_all(self): + self.clean_ref() + self.clean_contextref() + class BankStylesTimestepEmbedSequential: def __init__(self): + # ref self.var_bank = [] self.mean_bank = [] self.style_cfgs = [] self.cn_idx: list[int] = [] + # cref + self.c_var_bank = [] + self.c_mean_bank = [] + self.c_style_cfgs = [] + self.c_cn_idx: list[int] = [] - def get_avg_var_bank(self): - return sum(self.var_bank) / float(len(self.var_bank)) + def get_var_bank(self, ignore_contextref=False): + if ignore_contextref: + return self.var_bank + return self.var_bank + self.c_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 get_mean_bank(self, ignore_contextref=False): + if ignore_contextref: + return self.mean_bank + return self.mean_bank + self.c_mean_bank - def clean(self): + def get_style_cfgs(self, ignore_contextref=False): + if ignore_contextref: + return self.style_cfgs + return self.style_cfgs + self.c_style_cfgs + + def get_cn_idx(self, ignore_contextref=False): + if ignore_contextref: + return self.cn_idx + return self.cn_idx + self.c_cn_idx + + def clean_ref(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.mean_bank = [] + self.var_bank = [] + self.style_cfgs = [] self.cn_idx = [] + def clean_contextref(self): + del self.c_var_bank + del self.c_mean_bank + del self.c_style_cfgs + del self.c_cn_idx + self.c_var_bank = [] + self.c_mean_bank = [] + self.c_style_cfgs = [] + self.c_cn_idx = [] + + def clean_all(self): + self.clean_ref() + self.clean_contextref() + class InjectionBasicTransformerBlockHolder: def __init__(self, block: BasicTransformerBlock, idx=None): @@ -322,8 +412,14 @@ class InjectionBasicTransformerBlockHolder: else: block.forward = self.original_forward - def clean(self): - self.bank_styles.clean() + def clean_ref(self): + self.bank_styles.clean_ref() + + def clean_contextref(self): + self.bank_styles.clean_contextref() + + def clean_all(self): + self.bank_styles.clean_all() class InjectionTimestepEmbedSequentialHolder: @@ -339,8 +435,14 @@ class InjectionTimestepEmbedSequentialHolder: def restore(self, block: openaimodel.TimestepEmbedSequential): block.forward = self.original_forward - def clean(self): - self.bank_styles.clean() + def clean_ref(self): + self.bank_styles.clean_ref() + + def clean_contextref(self): + self.bank_styles.clean_contextref() + + def clean_all(self): + self.bank_styles.clean_all() class ReferenceInjections: @@ -349,20 +451,44 @@ class ReferenceInjections: self.gn_modules = gn_modules if gn_modules else [] self.diffusion_model_orig_forward: Callable = None - def clean_module_mem(self): + def clean_ref_module_mem(self): for attn_module in self.attn_modules: try: - attn_module.injection_holder.clean() + attn_module.injection_holder.clean_ref() except Exception: pass for gn_module in self.gn_modules: try: - gn_module.injection_holder.clean() + gn_module.injection_holder.clean_ref() + except Exception: + pass + + def clean_contextref_module_mem(self): + for attn_module in self.attn_modules: + try: + attn_module.injection_holder.clean_contextref() + except Exception: + pass + for gn_module in self.gn_modules: + try: + gn_module.injection_holder.clean_contextref() + except Exception: + pass + + def clean_all_module_mem(self): + for attn_module in self.attn_modules: + try: + attn_module.injection_holder.clean_all() + except Exception: + pass + for gn_module in self.gn_modules: + try: + gn_module.injection_holder.clean_all() except Exception: pass def cleanup(self): - self.clean_module_mem() + self.clean_all_module_mem() del self.attn_modules self.attn_modules = [] del self.gn_modules @@ -370,13 +496,262 @@ class ReferenceInjections: self.diffusion_model_orig_forward = None +def HACK_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", {}) + # 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) + + context_ref = ref_controlnets[0] + if transformer_options[CONTEXTREF_ATTN_MACHINE_STATE] in [MachineState.WRITE, MachineState.OFF]: + reference_injections.clean_module_mem() + transformer_options[REF_ATTN_MACHINE_STATE] = transformer_options[CONTEXTREF_ATTN_MACHINE_STATE] + transformer_options[REF_ADAIN_MACHINE_STATE] = transformer_options[CONTEXTREF_ADAIN_MACHINE_STATE] + transformer_options[REF_ATTN_CONTROL_LIST] = [context_ref] + transformer_options[REF_ADAIN_CONTROL_LIST] = [context_ref] + return reference_injections.diffusion_model_orig_forward(x, *args, **kwargs) + + # 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 + + + + 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 + + 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) + transformer_options: dict[str] = kwargs.get("transformer_options", {}) + # NOTE: adds support for both ReferenceCN and ContextRef, so need to track them separately + # get ReferenceAdvanced objects + ref_controlnets: list[ReferenceAdvanced] = transformer_options.get(REF_CONTROL_LIST_ALL, []) + context_controlnets: list[ReferenceAdvanced] = transformer_options.get(CONTEXTREF_CONTROL_LIST_ALL, []) + # discard any controlnets that should not run + ref_controlnets = [z for z in ref_controlnets if z.should_run()] + context_controlnets = [z for z in context_controlnets if z.should_run()] + # if nothing related to reference controlnets, do nothing special + if len(ref_controlnets) == 0 and len(context_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, z in enumerate(indiv_conds) if z == 1] + transformer_options[REF_COND_IDXS] = [i for i, z in enumerate(indiv_conds) if z == 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) + context_attn_controlnets = [] + context_adain_controlnets = [] + for control in context_controlnets: + if ReferenceType.is_attn(control.ref_opts.reference_type): + context_attn_controlnets.append(control) + if ReferenceType.is_adain(control.ref_opts.reference_type): + context_adain_controlnets.append(control) + if len(adain_controlnets) > 0 or len(context_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) + + # if RefCN to be used, handle running diffusion with ref cond hints + if len(ref_controlnets) > 0: + for control in ref_controlnets: + read_attn_list = [] + write_attn_list = [] + read_adain_list = [] + write_adain_list = [] + + if ReferenceType.is_attn(control.ref_opts.reference_type): + write_attn_list.append(control) + if ReferenceType.is_adain(control.ref_opts.reference_type): + write_adain_list.append(control) + # apply lists + transformer_options[REF_READ_ATTN_CONTROL_LIST] = read_attn_list + transformer_options[REF_WRITE_ATTN_CONTROL_LIST] = write_attn_list + transformer_options[REF_READ_ADAIN_CONTROL_LIST] = read_adain_list + transformer_options[REF_WRITE_ADAIN_CONTROL_LIST] = write_adain_list + + orig_kwargs = kwargs + # disable other controlnets for this run, if specified + 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 + # prepare running diffusion for real now + read_attn_list = [] + write_attn_list = [] + read_adain_list = [] + write_adain_list = [] + + # add RefCNs to read lists + read_attn_list.extend(attn_controlnets) + read_adain_list.extend(adain_controlnets) + + # do contextref stuff, if needed + if len(context_controlnets) > 0: + # TODO: clean contextref stuff if attn writing or off + if transformer_options[CONTEXTREF_ATTN_MACHINE_STATE] in [MachineState.WRITE, MachineState.OFF]: + reference_injections.clean_contextref_module_mem() + ### add ContextRef to appropriate lists + # attn + if transformer_options[CONTEXTREF_ATTN_MACHINE_STATE] == MachineState.WRITE: + write_attn_list.extend(context_attn_controlnets) + elif transformer_options[CONTEXTREF_ATTN_MACHINE_STATE] == MachineState.READ: + read_attn_list.extend(context_attn_controlnets) + # adain + if transformer_options[CONTEXTREF_ADAIN_MACHINE_STATE] == MachineState.WRITE: + write_attn_list.extend(context_adain_controlnets) + elif transformer_options[CONTEXTREF_ADAIN_MACHINE_STATE] == MachineState.READ: + read_attn_list.extend(context_adain_controlnets) + # apply lists, containing both RefCN and ContextRef + transformer_options[REF_READ_ATTN_CONTROL_LIST] = read_attn_list + transformer_options[REF_WRITE_ATTN_CONTROL_LIST] = write_attn_list + transformer_options[REF_READ_ADAIN_CONTROL_LIST] = read_adain_list + transformer_options[REF_WRITE_ADAIN_CONTROL_LIST] = write_adain_list + + 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_ref_module_mem() + if len(adain_controlnets) > 0: + openaimodel.forward_timestep_embed = orig_forward_timestep_embed + + return forward_inject_UNetModel + + + +def ORIG_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", {}) # 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 @@ -476,16 +851,29 @@ def _forward_inject_BasicTransformerBlock(self: RefBasicTransformerBlock, x: Ten # Reference CN stuff uc_idx_mask = transformer_options.get(REF_UNCOND_IDXS, []) - c_idx_mask = transformer_options.get(REF_COND_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) + ref_write_cns: list[ReferenceAdvanced] = transformer_options.get(REF_WRITE_ATTN_CONTROL_LIST, []) + ref_read_cns: list[ReferenceAdvanced] = transformer_options.get(REF_READ_ATTN_CONTROL_LIST, []) + ignore_contextref_read = False # if writing to bank, should NOT be read in the same execution + + # if any refs to WRITE, save n and style_fidelity + if len(ref_write_cns) > 0: + cached_n = None + for refcn in ref_write_cns: + if refcn.ref_opts.attn_ref_weight > self.injection_holder.attn_weight: + if cached_n is None: + cached_n = n.detach().clone() + if refcn.is_context_ref: # store separately for RefCN and ContextRef + self.injection_holder.bank_styles.c_bank.append(cached_n) + self.injection_holder.bank_styles.c_style_cfgs.append(ref_write_cns[0].ref_opts.attn_style_fidelity) + self.injection_holder.bank_styles.c_cn_idx.append(ref_write_cns[0].order) + ignore_contextref_read = True + else: + self.injection_holder.bank_styles.bank.append(cached_n) + self.injection_holder.bank_styles.style_cfgs.append(ref_write_cns[0].ref_opts.attn_style_fidelity) + self.injection_holder.bank_styles.cn_idx.append(ref_write_cns[0].order) + del cached_n if "attn1_patch" in transformer_patches: patch = transformer_patches["attn1_patch"] @@ -510,20 +898,20 @@ def _forward_inject_BasicTransformerBlock(self: RefBasicTransformerBlock, x: Ten 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: + if len(ref_read_cns) > 0 and len(self.injection_holder.bank_styles.get_bank(ignore_contextref_read)) > 0: bank_styles = self.injection_holder.bank_styles - style_fidelity = bank_styles.get_avg_style_fidelity() - real_bank = bank_styles.bank.copy() + style_fidelity = bank_styles.get_avg_style_fidelity(ignore_contextref_read) + real_bank = bank_styles.get_bank(ignore_contextref_read).copy() cn_idx = 0 - for idx, order in enumerate(bank_styles.cn_idx): + for idx, order in enumerate(bank_styles.get_cn_idxs(ignore_contextref_read)): # make sure matching ref cn is selected - for i in range(cn_idx, len(ref_controlnets)): - if ref_controlnets[i].order == order: + for i in range(cn_idx, len(ref_read_cns)): + if ref_read_cns[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) + assert order == ref_read_cns[cn_idx].order + if ref_read_cns[cn_idx].any_attn_strength_to_apply(): + effective_strength = ref_read_cns[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, @@ -538,7 +926,7 @@ def _forward_inject_BasicTransformerBlock(self: RefBasicTransformerBlock, x: Ten 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() + bank_styles.clean_ref() else: context_attn1 = self.attn1.to_k(context_attn1) value_attn1 = self.attn1.to_v(value_attn1) @@ -546,22 +934,22 @@ def _forward_inject_BasicTransformerBlock(self: RefBasicTransformerBlock, x: Ten 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 len(ref_read_cns) > 0 and len(self.injection_holder.bank_styles.get_bank(ignore_contextref_read)) > 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() + style_fidelity = bank_styles.get_avg_style_fidelity(ignore_contextref_read) + real_bank = bank_styles.get_bank(ignore_contextref_read).copy() cn_idx = 0 - for idx, order in enumerate(bank_styles.cn_idx): + for idx, order in enumerate(bank_styles.get_cn_idxs(ignore_contextref_read)): # make sure matching ref cn is selected - for i in range(cn_idx, len(ref_controlnets)): - if ref_controlnets[i].order == order: + for i in range(cn_idx, len(ref_read_cns)): + if ref_read_cns[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) + assert order == ref_read_cns[cn_idx].order + if ref_read_cns[cn_idx].any_attn_strength_to_apply(): + effective_strength = ref_read_cns[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, @@ -574,7 +962,7 @@ def _forward_inject_BasicTransformerBlock(self: RefBasicTransformerBlock, x: Ten 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() + bank_styles.clean_ref() else: n = self.attn1(n, context=context_attn1, value=value_attn1) @@ -647,41 +1035,53 @@ def forward_timestep_embed_ref_inject_factory(orig_timestep_embed_inject_factory 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, []) + #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: + ref_write_cns: list[ReferenceAdvanced] = transformer_options.get(REF_WRITE_ADAIN_CONTROL_LIST, []) + ref_read_cns: list[ReferenceAdvanced] = transformer_options.get(REF_READ_ADAIN_CONTROL_LIST, []) + ignore_contextref_read = False # if writing to bank, should NOT be read in the same execution + + # if any refs to WRITE, save var, mean, and style_cfg + for refcn in ref_write_cns: + if refcn.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: + if refcn.is_context_ref: + ts.injection_holder.bank_styles.c_var_bank.append(var) + ts.injection_holder.bank_styles.c_mean_bank.append(mean) + ts.injection_holder.bank_styles.c_style_cfgs.append(refcn.ref_opts.adain_style_fidelity) + ts.injection_holder.bank_styles.c_cn_idx.append(refcn.order) + ignore_contextref_read = True + else: + 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(refcn.ref_opts.adain_style_fidelity) + ts.injection_holder.bank_styles.cn_idx.append(refcn.order) + + # if any refs to READ, do math with saved var, mean, and style_cfg + if len(ref_read_cns) > 0: + if len(ts.injection_holder.bank_styles.get_var_bank(ignore_contextref_read)) > 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): + real_style_cfgs = bank_styles.get_style_cfgs(ignore_contextref_read) + real_var_bank = bank_styles.get_var_bank(ignore_contextref_read) + real_mean_bank = bank_styles.get_mean_bank(ignore_contextref_read) + for idx, order in enumerate(bank_styles.get_cn_idx(ignore_contextref_read)): # make sure matching ref cn is selected - for i in range(cn_idx, len(ref_controlnets)): - if ref_controlnets[i].order == order: + for i in range(cn_idx, len(ref_read_cns)): + if ref_read_cns[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] + assert order == ref_read_cns[cn_idx].order + style_fidelity = real_style_cfgs[idx] + var_acc = real_var_bank[idx] + mean_acc = real_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) + if ref_read_cns[cn_idx].any_adain_strength_to_apply(): + effective_strength = ref_read_cns[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 @@ -691,7 +1091,7 @@ def forward_timestep_embed_ref_inject_factory(orig_timestep_embed_inject_factory 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() + ts.injection_holder.bank_styles.clean_ref() if y is None: y = x diff --git a/adv_control/sampling.py b/adv_control/sampling.py index 81f0407..bec3c44 100644 --- a/adv_control/sampling.py +++ b/adv_control/sampling.py @@ -11,24 +11,39 @@ from .control_reference import (ReferenceAdvanced, ReferenceInjections, RefBasicTransformerBlock, RefTimestepEmbedSequential, InjectionBasicTransformerBlockHolder, InjectionTimestepEmbedSequentialHolder, _forward_inject_BasicTransformerBlock, factory_forward_inject_UNetModel, - REF_CONTROL_LIST_ALL) + handle_context_ref_setup, + REF_CONTROL_LIST_ALL, CONTEXTREF_CLEAN_FUNC) from .control_lllite import (ControlLLLiteAdvanced) from .utils import torch_dfs def support_sliding_context_windows(model, positive, negative) -> tuple[bool, dict, dict]: - if not hasattr(model, "motion_injection_params"): - return False, positive, negative - motion_injection_params = getattr(model, "motion_injection_params") - context_options = getattr(motion_injection_params, "context_options") - if context_options.context_length is None: - return False, positive, negative # convert to advanced, with report if anything was actually modified modified, new_conds = convert_all_to_advanced([positive, negative]) positive, negative = new_conds return modified, positive, negative +def has_sliding_context_windows(model): + motion_injection_params = getattr(model, "motion_injection_params", None) + if motion_injection_params is None: + return False + context_options = getattr(motion_injection_params, "context_options") + return context_options.context_length is not None + + +def has_contextref_enabled(model): + motion_injection_params = getattr(model, "motion_injection_params", None) + if motion_injection_params is None: + return False + context_options = getattr(motion_injection_params, "context_options") + extras = getattr(context_options, "extras", None) + if extras is None: + return False + context_ref = getattr(extras, "context_ref", None) + return context_ref is not None + + def acn_sample_factory(orig_comfy_sample: Callable, is_custom=False) -> Callable: def get_refcn(control: ControlBase, order: int=-1): ref_set: set[ReferenceAdvanced] = set() @@ -59,13 +74,20 @@ def acn_sample_factory(orig_comfy_sample: Callable, is_custom=False) -> Callable # check if positive or negative conds contain ref cn positive = args[-3] negative = args[-2] - # if context options present, convert all CNs to Advanced if needed - controlnets_modified, positive, negative = support_sliding_context_windows(model, positive, negative) - if controlnets_modified: - args = list(args) - args[-3] = positive - args[-2] = negative - args = tuple(args) + # if context options present, perform some special actions that may be required + context_refs = [] + if has_sliding_context_windows(model): + model.model_options = model.model_options.copy() + model.model_options["transformer_options"] = model.model_options["transformer_options"].copy() + if has_contextref_enabled(model): + context_refs = handle_context_ref_setup(model.model_options["transformer_options"]) + # convert all CNs to Advanced if needed + controlnets_modified, positive, negative = support_sliding_context_windows(model, positive, negative) + if controlnets_modified: + args = list(args) + args[-3] = positive + args[-2] = negative + args = tuple(args) # look for Advanced ControlNets that will require intervention to work ref_set = set() lllite_dict: dict[ControlLLLiteAdvanced, None] = {} # dicts preserve insertion order since py3.7 @@ -88,7 +110,7 @@ def acn_sample_factory(orig_comfy_sample: Callable, is_custom=False) -> Callable for lll in lllite_list: lll.live_model_patches(model.model_options) # if no ref cn found, do original function immediately - if len(ref_set) == 0: + if len(ref_set) == 0 and len(context_refs) == 0: return orig_comfy_sample(model, *args, **kwargs) # otherwise, injection time try: @@ -158,6 +180,7 @@ def acn_sample_factory(orig_comfy_sample: Callable, is_custom=False) -> Callable 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) + new_model_options["transformer_options"][CONTEXTREF_CLEAN_FUNC] = reference_injections.clean_contextref_module_mem model.model_options = new_model_options # continue with original function return orig_comfy_sample(model, *args, **kwargs) @@ -167,14 +190,14 @@ def acn_sample_factory(orig_comfy_sample: Callable, is_custom=False) -> Callable attn_modules: list[RefBasicTransformerBlock] = reference_injections.attn_modules for module in attn_modules: module.injection_holder.restore(module) - module.injection_holder.clean() + module.injection_holder.clean_all() 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() + module.injection_holder.clean_all() del module.injection_holder del gn_modules # restore diffusion_model forward function