import math from contextlib import contextmanager from dataclasses import dataclass from math import ceil from typing import Callable, Optional, Union, Any, Dict import numpy as np import torch from diffusers.models.cross_attention import AttnProcessor from ldm.models.diffusion.cross_attention_control import Arguments, \ remove_cross_attention_control, setup_cross_attention_control, Context, get_cross_attention_modules, \ CrossAttentionType, SwapCrossAttnContext from ldm.models.diffusion.cross_attention_map_saving import AttentionMapSaver @dataclass(frozen=True) class ThresholdSettings: threshold: float warmup: float class InvokeAIDiffuserComponent: ''' The aim of this component is to provide a single place for code that can be applied identically to all InvokeAI diffusion procedures. At the moment it includes the following features: * Cross attention control ("prompt2prompt") * Hybrid conditioning (used for inpainting) ''' debug_thresholding = False @dataclass class ExtraConditioningInfo: tokens_count_including_eos_bos: int cross_attention_control_args: Optional[Arguments] = None @property def wants_cross_attention_control(self): return self.cross_attention_control_args is not None def __init__(self, model, model_forward_callback: Callable[[torch.Tensor, torch.Tensor, torch.Tensor, Optional[dict[str,Any]]], torch.Tensor], is_running_diffusers: bool=False, ): """ :param model: the unet model to pass through to cross attention control :param model_forward_callback: a lambda with arguments (x, sigma, conditioning_to_apply). will be called repeatedly. most likely, this should simply call model.forward(x, sigma, conditioning) """ self.conditioning = None self.model = model self.is_running_diffusers = is_running_diffusers self.model_forward_callback = model_forward_callback self.cross_attention_control_context = None @contextmanager def custom_attention_context(self, extra_conditioning_info: Optional[ExtraConditioningInfo], step_count: int): do_swap = extra_conditioning_info is not None and extra_conditioning_info.wants_cross_attention_control old_attn_processor = None if do_swap: old_attn_processor = self.setup_cross_attention_control(extra_conditioning_info, step_count=step_count) try: yield None finally: if old_attn_processor is not None: self.remove_cross_attention_control(old_attn_processor) # TODO resuscitate attention map saving #self.remove_attention_map_saving() def setup_cross_attention_control(self, conditioning: ExtraConditioningInfo, step_count: int) -> Dict[str, AttnProcessor]: """ setup cross attention .swap control. for diffusers this replaces the attention processor, so the previous attention processor is returned so that the caller can restore it later. """ self.conditioning = conditioning self.cross_attention_control_context = Context( arguments=self.conditioning.cross_attention_control_args, step_count=step_count ) return setup_cross_attention_control(self.model, self.cross_attention_control_context, is_running_diffusers=self.is_running_diffusers) def remove_cross_attention_control(self, restore_attention_processor: Optional['AttnProcessor']=None): self.conditioning = None self.cross_attention_control_context = None remove_cross_attention_control(self.model, is_running_diffusers=self.is_running_diffusers, restore_attention_processor=restore_attention_processor) def setup_attention_map_saving(self, saver: AttentionMapSaver): def callback(slice, dim, offset, slice_size, key): if dim is not None: # sliced tokens attention map saving is not implemented return saver.add_attention_maps(slice, key) tokens_cross_attention_modules = get_cross_attention_modules(self.model, CrossAttentionType.TOKENS) for identifier, module in tokens_cross_attention_modules: key = ('down' if identifier.startswith('down') else 'up' if identifier.startswith('up') else 'mid') module.set_attention_slice_calculated_callback( lambda slice, dim, offset, slice_size, key=key: callback(slice, dim, offset, slice_size, key)) def remove_attention_map_saving(self): tokens_cross_attention_modules = get_cross_attention_modules(self.model, CrossAttentionType.TOKENS) for _, module in tokens_cross_attention_modules: module.set_attention_slice_calculated_callback(None) def do_diffusion_step(self, x: torch.Tensor, sigma: torch.Tensor, unconditioning: Union[torch.Tensor,dict], conditioning: Union[torch.Tensor,dict], unconditional_guidance_scale: float, step_index: Optional[int]=None, total_step_count: Optional[int]=None, threshold: Optional[ThresholdSettings]=None, ): """ :param x: current latents :param sigma: aka t, passed to the internal model to control how much denoising will occur :param unconditioning: embeddings for unconditioned output. for hybrid conditioning this is a dict of tensors [B x 77 x 768], otherwise a single tensor [B x 77 x 768] :param conditioning: embeddings for conditioned output. for hybrid conditioning this is a dict of tensors [B x 77 x 768], otherwise a single tensor [B x 77 x 768] :param unconditional_guidance_scale: aka CFG scale, controls how much effect the conditioning tensor has :param step_index: counts upwards from 0 to (step_count-1) (as passed to setup_cross_attention_control, if using). May be called multiple times for a single step, therefore do not assume that its value will monotically increase. If None, will be estimated by comparing sigma against self.model.sigmas . :param threshold: threshold to apply after each step :return: the new latents after applying the model to x using unscaled unconditioning and CFG-scaled conditioning. """ cross_attention_control_types_to_do = [] context: Context = self.cross_attention_control_context if self.cross_attention_control_context is not None: if step_index is not None and total_step_count is not None: # 🧨diffusers codepath percent_through = step_index / total_step_count # will never reach 1.0 - this is deliberate else: # legacy compvis codepath # TODO remove when compvis codepath support is dropped if step_index is None and sigma is None: raise ValueError(f"Either step_index or sigma is required when doing cross attention control, but both are None.") percent_through = self.estimate_percent_through(step_index, sigma) cross_attention_control_types_to_do = context.get_active_cross_attention_control_types_for_step(percent_through) wants_cross_attention_control = (len(cross_attention_control_types_to_do) > 0) wants_hybrid_conditioning = isinstance(conditioning, dict) if wants_hybrid_conditioning: unconditioned_next_x, conditioned_next_x = self.apply_hybrid_conditioning(x, sigma, unconditioning, conditioning) elif wants_cross_attention_control: unconditioned_next_x, conditioned_next_x = self.apply_cross_attention_controlled_conditioning(x, sigma, unconditioning, conditioning, cross_attention_control_types_to_do) else: unconditioned_next_x, conditioned_next_x = self.apply_standard_conditioning(x, sigma, unconditioning, conditioning) combined_next_x = self._combine(unconditioned_next_x, conditioned_next_x, unconditional_guidance_scale) if threshold: combined_next_x = self._threshold(threshold.threshold, threshold.warmup, combined_next_x, sigma) return combined_next_x # methods below are called from do_diffusion_step and should be considered private to this class. def apply_standard_conditioning(self, x, sigma, unconditioning, conditioning): # fast batched path x_twice = torch.cat([x] * 2) sigma_twice = torch.cat([sigma] * 2) both_conditionings = torch.cat([unconditioning, conditioning]) both_results = self.model_forward_callback(x_twice, sigma_twice, both_conditionings) unconditioned_next_x, conditioned_next_x = both_results.chunk(2) if conditioned_next_x.device.type == 'mps': # prevent a result filled with zeros. seems to be a torch bug. conditioned_next_x = conditioned_next_x.clone() return unconditioned_next_x, conditioned_next_x def apply_hybrid_conditioning(self, x, sigma, unconditioning, conditioning): assert isinstance(conditioning, dict) assert isinstance(unconditioning, dict) x_twice = torch.cat([x] * 2) sigma_twice = torch.cat([sigma] * 2) both_conditionings = dict() for k in conditioning: if isinstance(conditioning[k], list): both_conditionings[k] = [ torch.cat([unconditioning[k][i], conditioning[k][i]]) for i in range(len(conditioning[k])) ] else: both_conditionings[k] = torch.cat([unconditioning[k], conditioning[k]]) unconditioned_next_x, conditioned_next_x = self.model_forward_callback(x_twice, sigma_twice, both_conditionings).chunk(2) return unconditioned_next_x, conditioned_next_x def apply_cross_attention_controlled_conditioning(self, x: torch.Tensor, sigma, unconditioning, conditioning, cross_attention_control_types_to_do): if self.is_running_diffusers: return self.apply_cross_attention_controlled_conditioning__diffusers(x, sigma, unconditioning, conditioning, cross_attention_control_types_to_do) else: return self.apply_cross_attention_controlled_conditioning__compvis(x, sigma, unconditioning, conditioning, cross_attention_control_types_to_do) def apply_cross_attention_controlled_conditioning__diffusers(self, x: torch.Tensor, sigma, unconditioning, conditioning, cross_attention_control_types_to_do): context: Context = self.cross_attention_control_context cross_attn_processor_context = SwapCrossAttnContext(modified_text_embeddings=context.arguments.edited_conditioning, index_map=context.cross_attention_index_map, mask=context.cross_attention_mask, cross_attention_types_to_do=[]) # no cross attention for unconditioning (negative prompt) unconditioned_next_x = self.model_forward_callback(x, sigma, unconditioning, {"swap_cross_attn_context": cross_attn_processor_context}) # do requested cross attention types for conditioning (positive prompt) cross_attn_processor_context.cross_attention_types_to_do = cross_attention_control_types_to_do conditioned_next_x = self.model_forward_callback(x, sigma, conditioning, {"swap_cross_attn_context": cross_attn_processor_context}) return unconditioned_next_x, conditioned_next_x def apply_cross_attention_controlled_conditioning__compvis(self, x:torch.Tensor, sigma, unconditioning, conditioning, cross_attention_control_types_to_do): # print('pct', percent_through, ': doing cross attention control on', cross_attention_control_types_to_do) # slower non-batched path (20% slower on mac MPS) # We are only interested in using attention maps for conditioned_next_x, but batching them with generation of # unconditioned_next_x causes attention maps to *also* be saved for the unconditioned_next_x. # This messes app their application later, due to mismatched shape of dim 0 (seems to be 16 for batched vs. 8) # (For the batched invocation the `wrangler` function gets attention tensor with shape[0]=16, # representing batched uncond + cond, but then when it comes to applying the saved attention, the # wrangler gets an attention tensor which only has shape[0]=8, representing just self.edited_conditionings.) # todo: give CrossAttentionControl's `wrangler` function more info so it can work with a batched call as well. context:Context = self.cross_attention_control_context try: unconditioned_next_x = self.model_forward_callback(x, sigma, unconditioning) # process x using the original prompt, saving the attention maps #print("saving attention maps for", cross_attention_control_types_to_do) for ca_type in cross_attention_control_types_to_do: context.request_save_attention_maps(ca_type) _ = self.model_forward_callback(x, sigma, conditioning) context.clear_requests(cleanup=False) # process x again, using the saved attention maps to control where self.edited_conditioning will be applied #print("applying saved attention maps for", cross_attention_control_types_to_do) for ca_type in cross_attention_control_types_to_do: context.request_apply_saved_attention_maps(ca_type) edited_conditioning = self.conditioning.cross_attention_control_args.edited_conditioning conditioned_next_x = self.model_forward_callback(x, sigma, edited_conditioning) context.clear_requests(cleanup=True) except: context.clear_requests(cleanup=True) raise return unconditioned_next_x, conditioned_next_x def _combine(self, unconditioned_next_x, conditioned_next_x, guidance_scale): # to scale how much effect conditioning has, calculate the changes it does and then scale that scaled_delta = (conditioned_next_x - unconditioned_next_x) * guidance_scale combined_next_x = unconditioned_next_x + scaled_delta return combined_next_x def _threshold(self, threshold, warmup, latents: torch.Tensor, sigma) -> torch.Tensor: warmup_scale = (1 - sigma.item() / 1000) / warmup if warmup else math.inf if warmup_scale < 1: # This arithmetic based on https://github.com/invoke-ai/InvokeAI/pull/395 warming_threshold = 1 + (threshold - 1) * warmup_scale current_threshold = np.clip(warming_threshold, 1, threshold) else: current_threshold = threshold if current_threshold <= 0: return latents maxval = latents.max().item() minval = latents.min().item() scale = 0.7 # default value from #395 if self.debug_thresholding: std, mean = [i.item() for i in torch.std_mean(latents)] outside = torch.count_nonzero((latents < -current_threshold) | (latents > current_threshold)) print(f"\nThreshold: 𝜎={sigma.item()} threshold={current_threshold:.3f} (of {threshold:.3f})\n" f" | min, mean, max = {minval:.3f}, {mean:.3f}, {maxval:.3f}\tstd={std}\n" f" | {outside / latents.numel() * 100:.2f}% values outside threshold") if maxval < current_threshold and minval > -current_threshold: return latents if maxval > current_threshold: maxval = np.clip(maxval * scale, 1, current_threshold) if minval < -current_threshold: minval = np.clip(minval * scale, -current_threshold, -1) if self.debug_thresholding: outside = torch.count_nonzero((latents < minval) | (latents > maxval)) print(f" | min, , max = {minval:.3f}, , {maxval:.3f}\t(scaled by {scale})\n" f" | {outside / latents.numel() * 100:.2f}% values will be clamped") return latents.clamp(minval, maxval) def estimate_percent_through(self, step_index, sigma): if step_index is not None and self.cross_attention_control_context is not None: # percent_through will never reach 1.0 (but this is intended) return float(step_index) / float(self.cross_attention_control_context.step_count) # find the best possible index of the current sigma in the sigma sequence smaller_sigmas = torch.nonzero(self.model.sigmas <= sigma) sigma_index = smaller_sigmas[-1].item() if smaller_sigmas.shape[0] > 0 else 0 # flip because sigmas[0] is for the fully denoised image # percent_through must be <1 return 1.0 - float(sigma_index + 1) / float(self.model.sigmas.shape[0]) # print('estimated percent_through', percent_through, 'from sigma', sigma.item()) # todo: make this work @classmethod def apply_conjunction(cls, x, t, forward_func, uc, c_or_weighted_c_list, global_guidance_scale): x_in = torch.cat([x] * 2) t_in = torch.cat([t] * 2) # aka sigmas deltas = None uncond_latents = None weighted_cond_list = c_or_weighted_c_list if type(c_or_weighted_c_list) is list else [(c_or_weighted_c_list, 1)] # below is fugly omg num_actual_conditionings = len(c_or_weighted_c_list) conditionings = [uc] + [c for c,weight in weighted_cond_list] weights = [1] + [weight for c,weight in weighted_cond_list] chunk_count = ceil(len(conditionings)/2) deltas = None for chunk_index in range(chunk_count): offset = chunk_index*2 chunk_size = min(2, len(conditionings)-offset) if chunk_size == 1: c_in = conditionings[offset] latents_a = forward_func(x_in[:-1], t_in[:-1], c_in) latents_b = None else: c_in = torch.cat(conditionings[offset:offset+2]) latents_a, latents_b = forward_func(x_in, t_in, c_in).chunk(2) # first chunk is guaranteed to be 2 entries: uncond_latents + first conditioining if chunk_index == 0: uncond_latents = latents_a deltas = latents_b - uncond_latents else: deltas = torch.cat((deltas, latents_a - uncond_latents)) if latents_b is not None: deltas = torch.cat((deltas, latents_b - uncond_latents)) # merge the weighted deltas together into a single merged delta per_delta_weights = torch.tensor(weights[1:], dtype=deltas.dtype, device=deltas.device) normalize = False if normalize: per_delta_weights /= torch.sum(per_delta_weights) reshaped_weights = per_delta_weights.reshape(per_delta_weights.shape + (1, 1, 1)) deltas_merged = torch.sum(deltas * reshaped_weights, dim=0, keepdim=True) # old_return_value = super().forward(x, sigma, uncond, cond, cond_scale) # assert(0 == len(torch.nonzero(old_return_value - (uncond_latents + deltas_merged * cond_scale)))) return uncond_latents + deltas_merged * global_guidance_scale