# Copyright 2022 Lunar Ring. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import os, sys dp_git = "/home/lugo/git/" sys.path.append(os.path.join(dp_git,'garden4')) sys.path.append('util') import torch torch.backends.cudnn.benchmark = False import numpy as np import warnings warnings.filterwarnings('ignore') import time import subprocess import warnings import torch from tqdm.auto import tqdm from diffusers import StableDiffusionInpaintPipeline from diffusers import StableDiffusionPipeline from diffusers.schedulers import DDIMScheduler from PIL import Image import matplotlib.pyplot as plt import torch from movie_util import MovieSaver import datetime from typing import Callable, List, Optional, Union import inspect from threading import Thread torch.set_grad_enabled(False) #%% class LatentBlending(): def __init__( self, pipe: Union[StableDiffusionInpaintPipeline, StableDiffusionPipeline], device: str, height: int = 512, width: int = 512, num_inference_steps: int = 30, guidance_scale: float = 7.5, seed: int = 420, ): r""" Initializes the latent blending class. Args: device: str Compute device, e.g. cuda:0 height: int Height of the desired output image. The model was trained on 512. width: int Width of the desired output image. The model was trained on 512. num_inference_steps: int Number of diffusion steps. Larger values will take more compute time. guidance_scale: float Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598). `guidance_scale` is defined as `w` of equation 2. of [Imagen Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`, usually at the expense of lower image quality. seed: int Random seed. """ self.pipe = pipe self.device = device self.guidance_scale = guidance_scale self.num_inference_steps = num_inference_steps self.width = width self.height = height self.seed = seed # Inits self.check_asserts() self.init_mode() # Initialize vars self.prompt1 = "" self.prompt2 = "" self.tree_latents = [] self.tree_fracts = [] self.tree_status = [] self.tree_final_imgs = [] self.list_nmb_branches_prev = [] self.list_injection_idx_prev = [] self.text_embedding1 = None self.text_embedding2 = None self.stop_diffusion = False self.negative_prompt = None def check_asserts(self): r""" Runs Minimal set of sanity checks. """ assert self.pipe.scheduler._class_name == 'DDIMScheduler', 'Currently only the DDIMScheduler is supported.' def init_mode(self): r""" Automatically sets the mode of this class, depending on the supplied pipeline. """ if self.pipe._class_name == 'StableDiffusionInpaintPipeline': self.mask_empty = Image.fromarray(255*np.ones([self.width, self.height], dtype=np.uint8)) self.image_empty = Image.fromarray(np.zeros([self.width, self.height, 3], dtype=np.uint8)) self.image_source = None self.mask_image = None self.mode = 'inpaint' else: self.mode = 'standard' def init_inpainting( self, image_source: Union[Image.Image, np.ndarray] = None, mask_image: Union[Image.Image, np.ndarray] = None, init_empty: Optional[bool] = False, ): r""" Initializes inpainting with a source and maks image. Args: image_source: Union[Image.Image, np.ndarray] Source image onto which the mask will be applied. mask_image: Union[Image.Image, np.ndarray] Mask image, value = 0 will stay untouched, value = 255 subjet to diffusion init_empty: Optional[bool]: Initialize inpainting with an empty image and mask, effectively disabling inpainting. """ assert self.mode == 'inpaint', 'Initialize class with an inpainting pipeline!' if not init_empty: assert image_source is not None, "init_inpainting: you need to provide image_source" assert mask_image is not None, "init_inpainting: you need to provide mask_image" if type(image_source) == np.ndarray: image_source = Image.fromarray(image_source) self.image_source = image_source if type(mask_image) == np.ndarray: mask_image = Image.fromarray(mask_image) self.mask_image = mask_image else: self.mask_image = self.mask_empty self.image_source = self.image_empty def set_prompt1(self, prompt: str): r""" Sets the first prompt (for the first keyframe) including text embeddings. Args: prompt: str ABC trending on artstation painted by Greg Rutkowski """ prompt = prompt.replace("_", " ") self.prompt1 = prompt self.text_embedding1 = self.get_text_embeddings(self.prompt1) def set_prompt2(self, prompt: str): r""" Sets the second prompt (for the second keyframe) including text embeddings. Args: prompt: str XYZ trending on artstation painted by Greg Rutkowski """ prompt = prompt.replace("_", " ") self.prompt2 = prompt self.text_embedding2 = self.get_text_embeddings(self.prompt2) def run_transition( self, list_nmb_branches: List[int], list_injection_strength: List[float] = None, list_injection_idx: List[int] = None, recycle_img1: Optional[bool] = False, recycle_img2: Optional[bool] = False, fixed_seeds: Optional[List[int]] = None, ): r""" Returns a list of transition images using spherical latent blending. Args: list_nmb_branches: List[int]: list of the number of branches for each injection. list_injection_strength: List[float]: list of injection strengths within interval [0, 1), values need to be increasing. Alternatively you can direclty specify the list_injection_idx. list_injection_idx: List[int]: list of injection strengths within interval [0, 1), values need to be increasing. Alternatively you can specify the list_injection_strength. recycle_img1: Optional[bool]: Don't recompute the latents for the first keyframe (purely prompt1). Saves compute. recycle_img2: Optional[bool]: Don't recompute the latents for the second keyframe (purely prompt2). Saves compute. fixed_seeds: Optional[List[int)]: You can supply two seeds that are used for the first and second keyframe (prompt1 and prompt2). Otherwise random seeds will be taken. """ # Sanity checks first assert self.text_embedding1 is not None, 'Set the first text embedding with .set_prompt1(...) first' assert self.text_embedding2 is not None, 'Set the second text embedding with .set_prompt2(...) first' assert not((list_injection_strength is not None) and (list_injection_idx is not None)), "suppyl either list_injection_strength or list_injection_idx" if list_injection_strength is None: assert list_injection_idx is not None, "Supply either list_injection_idx or list_injection_strength" assert isinstance(list_injection_idx[0], int) or isinstance(list_injection_idx[0], np.int) , "Need to supply integers for list_injection_idx" if list_injection_idx is None: assert list_injection_strength is not None, "Supply either list_injection_idx or list_injection_strength" # Create the injection indexes list_injection_idx = [int(round(x*self.num_inference_steps)) for x in list_injection_strength] assert min(np.diff(list_injection_idx)) > 0, 'Injection idx needs to be increasing' if min(np.diff(list_injection_idx)) < 2: print("Warning: your injection spacing is very tight. consider increasing the distances") assert isinstance(list_injection_strength[1], np.floating) or isinstance(list_injection_strength[1], float), "Need to supply floats for list_injection_strength" # we are checking element 1 in list_injection_strength because "0" is an int... [0, 0.5] assert max(list_injection_idx) < self.num_inference_steps, "Decrease the injection index or strength" assert len(list_injection_idx) == len(list_nmb_branches), "Need to have same length" assert max(list_injection_idx) < self.num_inference_steps,"Injection index cannot happen after last diffusion step! Decrease list_injection_idx or list_injection_strength[-1]" if fixed_seeds is not None: if fixed_seeds == 'randomize': fixed_seeds = list(np.random.randint(0, 1000000, 2).astype(np.int32)) else: assert len(fixed_seeds)==2, "Supply a list with len = 2" # Process interruption variable self.stop_diffusion = False # Recycling? There are requirements if recycle_img1 or recycle_img2: if self.list_nmb_branches_prev == []: print("Warning. You want to recycle but there is nothing here. Disabling recycling.") recycle_img1 = False recycle_img2 = False elif self.list_nmb_branches_prev != list_nmb_branches: print("Warning. Cannot change list_nmb_branches if recycling latent. Disabling recycling.") recycle_img1 = False recycle_img2 = False elif self.list_injection_idx_prev != list_injection_idx: print("Warning. Cannot change list_nmb_branches if recycling latent. Disabling recycling.") recycle_img1 = False recycle_img2 = False # Make a backup for future reference self.list_nmb_branches_prev = list_nmb_branches self.list_injection_idx_prev = list_injection_idx # Auto inits list_injection_idx_ext = list_injection_idx[:] list_injection_idx_ext.append(self.num_inference_steps) # If injection at depth 0 not specified, we will start out with 2 branches if list_injection_idx_ext[0] != 0: list_injection_idx_ext.insert(0,0) list_nmb_branches.insert(0,2) assert list_nmb_branches[0] == 2, "Need to start with 2 branches. set list_nmb_branches[0]=2" # Pre-define entire branching tree structures if not recycle_img1 and not recycle_img2: self.tree_latents = [] self.tree_fracts = [] self.tree_status = [] self.tree_final_imgs = [None]*list_nmb_branches[-1] self.tree_final_imgs_timing = [0]*list_nmb_branches[-1] nmb_blocks_time = len(list_injection_idx_ext)-1 for t_block in range(nmb_blocks_time): nmb_branches = list_nmb_branches[t_block] list_fract_mixing_current = np.linspace(0, 1, nmb_branches) self.tree_fracts.append(list_fract_mixing_current) self.tree_latents.append([None]*nmb_branches) self.tree_status.append(['untouched']*nmb_branches) else: self.tree_final_imgs = [None]*list_nmb_branches[-1] nmb_blocks_time = len(list_injection_idx_ext)-1 for t_block in range(nmb_blocks_time): nmb_branches = list_nmb_branches[t_block] for idx_branch in range(nmb_branches): self.tree_status[t_block][idx_branch] = 'untouched' if recycle_img1: self.tree_status[t_block][0] = 'computed' self.tree_final_imgs[0] = self.latent2image(self.tree_latents[-1][0][-1]) self.tree_final_imgs_timing[0] = 0 if recycle_img2: self.tree_status[t_block][-1] = 'computed' self.tree_final_imgs[-1] = self.latent2image(self.tree_latents[-1][-1][-1]) self.tree_final_imgs_timing[-1] = 0 # setup compute order: goal: try to get last branch computed asap. # first compute the right keyframe. needs to be there in any case list_compute = [] list_local_stem = [] for t_block in range(nmb_blocks_time - 1, -1, -1): if self.tree_status[t_block][0] == 'untouched': self.tree_status[t_block][0] = 'prefetched' list_local_stem.append([t_block, 0]) list_compute.extend(list_local_stem[::-1]) # setup compute order: start from last leafs (the final transition images) and work way down. what parents do they need? for idx_leaf in range(1, list_nmb_branches[-1]): list_local_stem = [] t_block = nmb_blocks_time - 1 t_block_prev = t_block - 1 self.tree_status[t_block][idx_leaf] = 'prefetched' list_local_stem.append([t_block, idx_leaf]) idx_leaf_deep = idx_leaf for t_block in range(nmb_blocks_time-1, 0, -1): t_block_prev = t_block - 1 fract_mixing = self.tree_fracts[t_block][idx_leaf_deep] list_fract_mixing_prev = self.tree_fracts[t_block_prev] b_parent1, b_parent2 = get_closest_idx(fract_mixing, list_fract_mixing_prev) assert self.tree_status[t_block_prev][b_parent1] != 'untouched', 'Branch destruction??? This should never happen!' if self.tree_status[t_block_prev][b_parent2] == 'untouched': self.tree_status[t_block_prev][b_parent2] = 'prefetched' list_local_stem.append([t_block_prev, b_parent2]) idx_leaf_deep = b_parent2 list_compute.extend(list_local_stem[::-1]) # Diffusion computations start here time_start = time.time() for t_block, idx_branch in tqdm(list_compute, desc="computing transition"): if self.stop_diffusion: print("run_transition: process interrupted") return self.tree_final_imgs # print(f"computing t_block {t_block} idx_branch {idx_branch}") idx_stop = list_injection_idx_ext[t_block+1] fract_mixing = self.tree_fracts[t_block][idx_branch] text_embeddings_mix = interpolate_linear(self.text_embedding1, self.text_embedding2, fract_mixing) if t_block == 0: if fixed_seeds is not None: if idx_branch == 0: self.set_seed(fixed_seeds[0]) elif idx_branch == list_nmb_branches[0] -1: self.set_seed(fixed_seeds[1]) list_latents = self.run_diffusion(text_embeddings_mix, idx_stop=idx_stop) else: # find parents latents b_parent1, b_parent2 = get_closest_idx(fract_mixing, self.tree_fracts[t_block-1]) latents1 = self.tree_latents[t_block-1][b_parent1][-1] if fract_mixing == 0: latents2 = latents1 else: latents2 = self.tree_latents[t_block-1][b_parent2][-1] idx_start = list_injection_idx_ext[t_block] fract_mixing_parental = (fract_mixing - self.tree_fracts[t_block-1][b_parent1]) / (self.tree_fracts[t_block-1][b_parent2] - self.tree_fracts[t_block-1][b_parent1]) latents_for_injection = interpolate_spherical(latents1, latents2, fract_mixing_parental) list_latents = self.run_diffusion(text_embeddings_mix, latents_for_injection, idx_start=idx_start, idx_stop=idx_stop) self.tree_latents[t_block][idx_branch] = list_latents self.tree_status[t_block][idx_branch] = 'computed' # Convert latents to image directly for the last t_block if t_block == nmb_blocks_time-1: self.tree_final_imgs[idx_branch] = self.latent2image(list_latents[-1]) self.tree_final_imgs_timing[idx_branch] = time.time() - time_start return self.tree_final_imgs def run_multi_transition( self, list_prompts: List[str], list_seeds: List[int] = None, list_nmb_branches: List[int] = None, list_injection_strength: List[float] = None, list_injection_idx: List[int] = None, ms: MovieSaver = None, fps: float = 24, duration_single_trans: float = 15, ): r""" Runs multiple transitions and stitches them together. You can supply the seeds for each prompt. Args: list_prompts: List[float]: list of the prompts. There will be a transition starting from the first to the last. list_seeds: List[int] = None: Random Seeds for each prompt. list_nmb_branches: List[int]: list of the number of branches for each injection. list_injection_strength: List[float]: list of injection strengths within interval [0, 1), values need to be increasing. Alternatively you can direclty specify the list_injection_idx. list_injection_idx: List[int]: list of injection strengths within interval [0, 1), values need to be increasing. Alternatively you can specify the list_injection_strength. ms: MovieSaver You need to spawn a moviesaver instance. fps: float: frames per second duration_single_trans: float: The duration of a single transition prompt[i] -> prompt[i+1]. The duration of your movie will be duration_single_trans * len(list_prompts) """ if list_seeds is None: list_seeds = list(np.random.randint(0, 10e10, len(list_prompts))) assert len(list_prompts) == len(list_seeds), "Supply the same number of prompts and seeds" for i in range(len(list_prompts)-1): print(f"Starting movie segment {i+1}/{len(list_prompts)-1}") if i==0: self.set_prompt1(list_prompts[i]) self.set_prompt2(list_prompts[i+1]) recycle_img1 = False else: self.swap_forward() self.set_prompt2(list_prompts[i+1]) recycle_img1 = True local_seeds = [list_seeds[i], list_seeds[i+1]] list_imgs = lb.run_transition(list_nmb_branches, list_injection_strength=list_injection_strength, list_injection_idx=list_injection_idx, recycle_img1=recycle_img1, fixed_seeds=local_seeds) list_imgs_interp = add_frames_linear_interp(list_imgs, fps, duration_single_trans) # Save movie frame for img in list_imgs_interp: ms.write_frame(img) ms.finalize() print("run_multi_transition: All completed.") @torch.no_grad() def run_diffusion( self, text_embeddings: torch.FloatTensor, latents_for_injection: torch.FloatTensor = None, idx_start: int = -1, idx_stop: int = -1, return_image: Optional[bool] = False ): r""" Wrapper function for run_diffusion_standard and run_diffusion_inpaint. Depending on the mode, the correct one will be executed. Args: text_embeddings: torch.FloatTensor Text embeddings used for diffusion latents_for_injection: torch.FloatTensor Latents that are used for injection idx_start: int Index of the diffusion process start and where the latents_for_injection are injected idx_stop: int Index of the diffusion process end. return_image: Optional[bool] Optionally return image directly """ if self.mode == 'standard': return self.run_diffusion_standard(text_embeddings, latents_for_injection=latents_for_injection, idx_start=idx_start, idx_stop=idx_stop, return_image=return_image) elif self.mode == 'inpaint': assert self.image_source is not None, "image_source is None. Please run init_inpainting first." assert self.mask_image is not None, "image_source is None. Please run init_inpainting first." return self.run_diffusion_inpaint(text_embeddings, latents_for_injection=latents_for_injection, idx_start=idx_start, idx_stop=idx_stop, return_image=return_image) @torch.no_grad() def run_diffusion_standard( self, text_embeddings: torch.FloatTensor, latents_for_injection: torch.FloatTensor = None, idx_start: int = -1, idx_stop: int = -1, return_image: Optional[bool] = False ): r""" Runs regular diffusion. Returns a list of latents that were computed. Adaptations allow to supply a) starting index for diffusion b) stopping index for diffusion c) latent representations that are injected at the starting index Furthermore the intermittent latents are collected and returned. Adapted from diffusers (https://github.com/huggingface/diffusers) Args: text_embeddings: torch.FloatTensor Text embeddings used for diffusion latents_for_injection: torch.FloatTensor Latents that are used for injection idx_start: int Index of the diffusion process start and where the latents_for_injection are injected idx_stop: int Index of the diffusion process end. return_image: Optional[bool] Optionally return image directly """ if latents_for_injection is None: do_inject_latents = False else: do_inject_latents = True generator = torch.Generator(device=self.device).manual_seed(int(self.seed)) batch_size = 1 height = self.height width = self.width num_inference_steps = self.num_inference_steps num_images_per_prompt = 1 do_classifier_free_guidance = True # duplicate text embeddings for each generation per prompt, using mps friendly method bs_embed, seq_len, _ = text_embeddings.shape text_embeddings = text_embeddings.repeat(1, num_images_per_prompt, 1) text_embeddings = text_embeddings.view(bs_embed * num_images_per_prompt, seq_len, -1) # set timesteps self.pipe.scheduler.set_timesteps(num_inference_steps) # Some schedulers like PNDM have timesteps as arrays # It's more optimized to move all timesteps to correct device beforehand timesteps_tensor = self.pipe.scheduler.timesteps.to(self.pipe.device) if not do_inject_latents: # get the initial random noise unless the user supplied it latents_shape = (batch_size * num_images_per_prompt, self.pipe.unet.in_channels, height // 8, width // 8) latents_dtype = text_embeddings.dtype latents = torch.randn(latents_shape, generator=generator, device=self.pipe.device, dtype=latents_dtype) # scale the initial noise by the standard deviation required by the scheduler latents = latents * self.pipe.scheduler.init_noise_sigma extra_step_kwargs = {} # collect latents list_latents_out = [] for i, t in enumerate(timesteps_tensor): if do_inject_latents: # Inject latent at right place if i < idx_start: continue elif i == idx_start: latents = latents_for_injection.clone() if i == idx_stop: return list_latents_out # expand the latents if we are doing classifier free guidance latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents latent_model_input = self.pipe.scheduler.scale_model_input(latent_model_input, t) # predict the noise residual noise_pred = self.pipe.unet(latent_model_input, t, encoder_hidden_states=text_embeddings).sample # perform guidance if do_classifier_free_guidance: noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond) # compute the previous noisy sample x_t -> x_t-1 latents = self.pipe.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample list_latents_out.append(latents.clone()) if return_image: return self.latent2image(latents) else: return list_latents_out @torch.no_grad() def run_diffusion_inpaint( self, text_embeddings: torch.FloatTensor, latents_for_injection: torch.FloatTensor = None, idx_start: int = -1, idx_stop: int = -1, return_image: Optional[bool] = False ): r""" Runs inpaint-based diffusion. Returns a list of latents that were computed. Adaptations allow to supply a) starting index for diffusion b) stopping index for diffusion c) latent representations that are injected at the starting index Furthermore the intermittent latents are collected and returned. Adapted from diffusers (https://github.com/huggingface/diffusers) Args: text_embeddings: torch.FloatTensor Text embeddings used for diffusion latents_for_injection: torch.FloatTensor Latents that are used for injection idx_start: int Index of the diffusion process start and where the latents_for_injection are injected idx_stop: int Index of the diffusion process end. return_image: Optional[bool] Optionally return image directly """ if latents_for_injection is None: do_inject_latents = False else: do_inject_latents = True generator = torch.Generator(device=self.device).manual_seed(int(self.seed)) batch_size = 1 height = self.height width = self.width num_inference_steps = self.num_inference_steps num_images_per_prompt = 1 do_classifier_free_guidance = True # prepare mask and masked_image mask, masked_image = self.prepare_mask_and_masked_image(self.image_source, self.mask_image) mask = mask.to(device=self.pipe.device, dtype=text_embeddings.dtype) masked_image = masked_image.to(device=self.pipe.device, dtype=text_embeddings.dtype) # resize the mask to latents shape as we concatenate the mask to the latents mask = torch.nn.functional.interpolate(mask, size=(height // 8, width // 8)) # encode the mask image into latents space so we can concatenate it to the latents masked_image_latents = self.pipe.vae.encode(masked_image).latent_dist.sample(generator=generator) masked_image_latents = 0.18215 * masked_image_latents # duplicate mask and masked_image_latents for each generation per prompt, using mps friendly method mask = mask.repeat(num_images_per_prompt, 1, 1, 1) masked_image_latents = masked_image_latents.repeat(num_images_per_prompt, 1, 1, 1) mask = torch.cat([mask] * 2) if do_classifier_free_guidance else mask masked_image_latents = ( torch.cat([masked_image_latents] * 2) if do_classifier_free_guidance else masked_image_latents ) num_channels_mask = mask.shape[1] num_channels_masked_image = masked_image_latents.shape[1] num_channels_latents = self.pipe.vae.config.latent_channels latents_shape = (batch_size * num_images_per_prompt, num_channels_latents, height // 8, width // 8) latents_dtype = text_embeddings.dtype latents = torch.randn(latents_shape, generator=generator, device=self.pipe.device, dtype=latents_dtype) latents = latents.to(self.pipe.device) # set timesteps self.pipe.scheduler.set_timesteps(num_inference_steps) timesteps_tensor = self.pipe.scheduler.timesteps.to(self.pipe.device) latents = latents * self.pipe.scheduler.init_noise_sigma extra_step_kwargs = {} # collect latents list_latents_out = [] for i, t in enumerate(timesteps_tensor): if do_inject_latents: # Inject latent at right place if i < idx_start: continue elif i == idx_start: latents = latents_for_injection.clone() if i == idx_stop: return list_latents_out # expand the latents if we are doing classifier free guidance latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents # concat latents, mask, masked_image_latents in the channel dimension latent_model_input = torch.cat([latent_model_input, mask, masked_image_latents], dim=1) latent_model_input = self.pipe.scheduler.scale_model_input(latent_model_input, t) # predict the noise residual noise_pred = self.pipe.unet(latent_model_input, t, encoder_hidden_states=text_embeddings).sample # perform guidance if do_classifier_free_guidance: noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond) # compute the previous noisy sample x_t -> x_t-1 latents = self.pipe.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample list_latents_out.append(latents.clone()) if return_image: return self.latent2image(latents) else: return list_latents_out @torch.no_grad() def latent2image( self, latents: torch.FloatTensor ): r""" Returns an image provided a latent representation from diffusion. Args: latents: torch.FloatTensor Result of the diffusion process. """ latents = 1 / 0.18215 * latents image = self.pipe.vae.decode(latents).sample image = (image / 2 + 0.5).clamp(0, 1) image = image.cpu().permute(0, 2, 3, 1).float().numpy() image = (image[0,:,:,:] * 255).astype(np.uint8) return image @torch.no_grad() def get_text_embeddings( self, prompt: str ): r""" Computes the text embeddings provided a string with a prompts. Adapted from diffusers (https://github.com/huggingface/diffusers) Args: prompt: str ABC trending on artstation painted by Old Greg. """ if self.negative_prompt is None: uncond_tokens = [""] else: if isinstance(self.negative_prompt, str): uncond_tokens = [self.negative_prompt] batch_size = 1 num_images_per_prompt = 1 do_classifier_free_guidance = True # get prompt text embeddings text_inputs = self.pipe.tokenizer( prompt, padding="max_length", max_length=self.pipe.tokenizer.model_max_length, return_tensors="pt", ) text_input_ids = text_inputs.input_ids # if text_input_ids.shape[-1] > self.pipe.tokenizer.modeLatentBlendingl_max_length: # removed_text = self.pipe.tokenizer.batch_decode(text_input_ids[:, self.pipe.tokenizer.model_max_length :]) # text_input_ids = text_input_ids[:, : self.pipe.tokenizer.model_max_length] text_embeddings = self.pipe.text_encoder(text_input_ids.to(self.pipe.device))[0] # duplicate text embeddings for each generation per prompt, using mps friendly method bs_embed, seq_len, _ = text_embeddings.shape text_embeddings = text_embeddings.repeat(1, num_images_per_prompt, 1) text_embeddings = text_embeddings.view(bs_embed * num_images_per_prompt, seq_len, -1) # get unconditional embeddings for classifier free guidance if do_classifier_free_guidance: max_length = text_input_ids.shape[-1] uncond_input = self.pipe.tokenizer( uncond_tokens, padding="max_length", max_length=max_length, truncation=True, return_tensors="pt", ) uncond_embeddings = self.pipe.text_encoder(uncond_input.input_ids.to(self.pipe.device))[0] seq_len = uncond_embeddings.shape[1] uncond_embeddings = uncond_embeddings.repeat(batch_size, num_images_per_prompt, 1) uncond_embeddings = uncond_embeddings.view(batch_size * num_images_per_prompt, seq_len, -1) text_embeddings = torch.cat([uncond_embeddings, text_embeddings]) return text_embeddings def prepare_mask_and_masked_image(self, image, mask): r""" Mask and image preparation for inpainting. Adapted from diffusers (https://github.com/huggingface/diffusers) Args: image: Source image mask: Mask image """ image = np.array(image.convert("RGB")) image = image[None].transpose(0, 3, 1, 2) image = torch.from_numpy(image).to(dtype=torch.float32) / 127.5 - 1.0 mask = np.array(mask.convert("L")) mask = mask.astype(np.float32) / 255.0 mask = mask[None, None] mask[mask < 0.5] = 0 mask[mask >= 0.5] = 1 mask = torch.from_numpy(mask) masked_image = image * (mask < 0.5) return mask, masked_image def randomize_seed(self): r""" Set a random seed for a fresh start. """ seed = np.random.randint(999999999) self.set_seed(seed) def set_seed(self, seed: int): r""" Set a the seed for a fresh start. """ self.seed = seed def swap_forward(self): r""" Moves over keyframe two -> keyframe one. Useful for making a sequence of transitions. """ # Move over all latents for t_block in range(len(self.tree_latents)): self.tree_latents[t_block][0] = self.tree_latents[t_block][-1] # Move over prompts and text embeddings self.prompt1 = self.prompt2 self.text_embedding1 = self.text_embedding2 # Final cleanup for extra sanity self.tree_final_imgs = [] # Auxiliary functions def get_closest_idx( fract_mixing: float, list_fract_mixing_prev: List[float], ): r""" Helper function to retrieve the parents for any given mixing. Example: fract_mixing = 0.4 and list_fract_mixing_prev = [0, 0.3, 0.6, 1.0] Will return the two closest values from list_fract_mixing_prev, i.e. [1, 2] """ pdist = fract_mixing - np.asarray(list_fract_mixing_prev) pdist_pos = pdist.copy() pdist_pos[pdist_pos<0] = np.inf b_parent1 = np.argmin(pdist_pos) pdist_neg = -pdist.copy() pdist_neg[pdist_neg<=0] = np.inf b_parent2= np.argmin(pdist_neg) if b_parent1 > b_parent2: tmp = b_parent2 b_parent2 = b_parent1 b_parent1 = tmp return b_parent1, b_parent2 @torch.no_grad() def interpolate_spherical(p0, p1, fract_mixing: float): r""" Helper function to correctly mix two random variables using spherical interpolation. See https://en.wikipedia.org/wiki/Slerp The function will always cast up to float64 for sake of extra precision. Args: p0: First tensor for interpolation p1: Second tensor for interpolation fract_mixing: float Mixing coefficient of interval [0, 1]. 0 will return in p0 1 will return in p1 0.x will return a mix between both preserving angular velocity. """ if p0.dtype == torch.float16: recast_to = 'fp16' else: recast_to = 'fp32' p0 = p0.double() p1 = p1.double() norm = torch.linalg.norm(p0) * torch.linalg.norm(p1) epsilon = 1e-7 dot = torch.sum(p0 * p1) / norm dot = dot.clamp(-1+epsilon, 1-epsilon) theta_0 = torch.arccos(dot) sin_theta_0 = torch.sin(theta_0) theta_t = theta_0 * fract_mixing s0 = torch.sin(theta_0 - theta_t) / sin_theta_0 s1 = torch.sin(theta_t) / sin_theta_0 interp = p0*s0 + p1*s1 if recast_to == 'fp16': interp = interp.half() elif recast_to == 'fp32': interp = interp.float() return interp def interpolate_linear(p0, p1, fract_mixing): r""" Helper function to mix two variables using standard linear interpolation. Args: p0: First tensor / np.ndarray for interpolation p1: Second tensor / np.ndarray for interpolation fract_mixing: float Mixing coefficient of interval [0, 1]. 0 will return in p0 1 will return in p1 0.x will return a linear mix between both. """ reconvert_uint8 = False if type(p0) is np.ndarray and p0.dtype == 'uint8': reconvert_uint8 = True p0 = p0.astype(np.float64) if type(p1) is np.ndarray and p1.dtype == 'uint8': reconvert_uint8 = True p1 = p1.astype(np.float64) interp = (1-fract_mixing) * p0 + fract_mixing * p1 if reconvert_uint8: interp = np.clip(interp, 0, 255).astype(np.uint8) return interp def add_frames_linear_interp( list_imgs: List[np.ndarray], fps_target: Union[float, int] = None, duration_target: Union[float, int] = None, nmb_frames_target: int=None, ): r""" Helper function to cheaply increase the number of frames given a list of images, by virtue of standard linear interpolation. The number of inserted frames will be automatically adjusted so that the total of number of frames can be fixed precisely, using a random shuffling technique. The function allows 1:1 comparisons between transitions as videos. Args: list_imgs: List[np.ndarray) List of images, between each image new frames will be inserted via linear interpolation. fps_target: OptionA: specify here the desired frames per second. duration_target: OptionA: specify here the desired duration of the transition in seconds. nmb_frames_target: OptionB: directly fix the total number of frames of the output. """ # Sanity if nmb_frames_target is not None and fps_target is not None: raise ValueError("You cannot specify both fps_target and nmb_frames_target") if fps_target is None: assert nmb_frames_target is not None, "Either specify nmb_frames_target or nmb_frames_target" if nmb_frames_target is None: assert fps_target is not None, "Either specify duration_target and fps_target OR nmb_frames_target" assert duration_target is not None, "Either specify duration_target and fps_target OR nmb_frames_target" nmb_frames_target = fps_target*duration_target # Get number of frames that are missing nmb_frames_diff = len(list_imgs)-1 nmb_frames_missing = nmb_frames_target - nmb_frames_diff - 1 if nmb_frames_missing < 1: return list_imgs list_imgs_float = [img.astype(np.float32) for img in list_imgs] # Distribute missing frames, append nmb_frames_to_insert(i) frames for each frame mean_nmb_frames_insert = nmb_frames_missing/nmb_frames_diff constfact = np.floor(mean_nmb_frames_insert) remainder_x = 1-(mean_nmb_frames_insert - constfact) nmb_iter = 0 while True: nmb_frames_to_insert = np.random.rand(nmb_frames_diff) nmb_frames_to_insert[nmb_frames_to_insert<=remainder_x] = 0 nmb_frames_to_insert[nmb_frames_to_insert>remainder_x] = 1 nmb_frames_to_insert += constfact if np.sum(nmb_frames_to_insert) == nmb_frames_missing: break nmb_iter += 1 if nmb_iter > 100000: print("add_frames_linear_interp: issue with inserting the right number of frames") break nmb_frames_to_insert = nmb_frames_to_insert.astype(np.int32) list_imgs_interp = [] for i in tqdm(range(len(list_imgs_float)-1), desc="STAGE linear interp"): img0 = list_imgs_float[i] img1 = list_imgs_float[i+1] list_imgs_interp.append(img0.astype(np.uint8)) list_fracts_linblend = np.linspace(0, 1, nmb_frames_to_insert[i]+2)[1:-1] for fract_linblend in list_fracts_linblend: img_blend = interpolate_linear(img0, img1, fract_linblend).astype(np.uint8) list_imgs_interp.append(img_blend.astype(np.uint8)) if i==len(list_imgs_float)-2: list_imgs_interp.append(img1.astype(np.uint8)) return list_imgs_interp def get_time(resolution=None): """ Helper function returning an nicely formatted time string, e.g. 221117_1620 """ if resolution==None: resolution="second" if resolution == "day": t = time.strftime('%y%m%d', time.localtime()) elif resolution == "minute": t = time.strftime('%y%m%d_%H%M', time.localtime()) elif resolution == "second": t = time.strftime('%y%m%d_%H%M%S', time.localtime()) elif resolution == "millisecond": t = time.strftime('%y%m%d_%H%M%S', time.localtime()) t += "_" t += str("{:03d}".format(int(int(datetime.utcnow().strftime('%f'))/1000))) else: raise ValueError("bad resolution provided: %s" %resolution) return t def get_branching( quality: str = 'medium', deepth_strength: float = 0.65, nmb_frames: int = 360, nmb_mindist: int = 3, ): r""" Helper function to set up the branching structure automatically. Args: quality: str Determines how many diffusion steps are being made + how many branches in total. Choose: fast, medium, high, ultra deepth_strength: float = 0.65, Determines how deep the first injection will happen. Deeper injections will cause (unwanted) formation of new structures, more shallow values will go into alpha-blendy land. nmb_frames: int = 360, total number of frames nmb_mindist: int = 3 minimum distance in terms of diffusion iteratinos between subsequent injections """ if quality == 'lowest': num_inference_steps = 12 nmb_branches_final = 5 elif quality == 'low': num_inference_steps = 15 nmb_branches_final = nmb_frames//16 elif quality == 'medium': num_inference_steps = 30 nmb_branches_final = nmb_frames//8 elif quality == 'high': num_inference_steps = 60 nmb_branches_final = nmb_frames//4 elif quality == 'ultra': num_inference_steps = 100 nmb_branches_final = nmb_frames//2 else: raise ValueError("quality = '{quality}' not supported") idx_injection_first = int(np.round(num_inference_steps*deepth_strength)) idx_injection_last = num_inference_steps - 3 nmb_injections = int(np.floor(num_inference_steps/5)) - 1 list_injection_idx = [0] list_injection_idx.extend(np.linspace(idx_injection_first, idx_injection_last, nmb_injections).astype(int)) list_nmb_branches = np.round(np.logspace(np.log10(2), np.log10(nmb_branches_final), nmb_injections+1)).astype(int) # Cleanup. There should be at least 3 diffusion steps between each injection list_injection_idx_clean = [list_injection_idx[0]] list_nmb_branches_clean = [list_nmb_branches[0]] idx_last_check = 0 for i in range(len(list_injection_idx)-1): if list_injection_idx[i+1] - list_injection_idx_clean[idx_last_check] >= nmb_mindist: list_injection_idx_clean.append(list_injection_idx[i+1]) list_nmb_branches_clean.append(list_nmb_branches[i+1]) idx_last_check +=1 list_injection_idx_clean = [int(l) for l in list_injection_idx_clean] list_nmb_branches_clean = [int(l) for l in list_nmb_branches_clean] print(f"num_inference_steps: {num_inference_steps}") print(f"list_injection_idx: {list_injection_idx_clean}") print(f"list_nmb_branches: {list_nmb_branches_clean}") return num_inference_steps, list_injection_idx_clean, list_nmb_branches_clean #%% le main if __name__ == "__main__": device = "cuda:0" model_path = "../stable_diffusion_models/stable-diffusion-v1-5" scheduler = DDIMScheduler(beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear", clip_sample=False, set_alpha_to_one=False) pipe = StableDiffusionPipeline.from_pretrained( model_path, revision="fp16", torch_dtype=torch.float16, scheduler=scheduler, use_auth_token=True ) pipe = pipe.to(device) num_inference_steps = 20 # Number of diffusion interations list_nmb_branches = [2, 3, 10, 24] # Branching structure: how many branches list_injection_strength = [0.0, 0.6, 0.8, 0.9] # Branching structure: how deep is the blending width = 512 height = 512 guidance_scale = 5 fixed_seeds = [993621550, 280335986] lb = LatentBlending(pipe, device, height, width, num_inference_steps, guidance_scale) lb.negative_prompt = 'text, letters' prompt1 = "photo of a beautiful newspaper covered in white flowers, ambient light, very detailed, magic" prompt2 = "photo of an eerie statue surrounded by ferns and vines, analog photograph kodak portra, mystical ambience, incredible detail" lb.set_prompt1(prompt1) lb.set_prompt2(prompt2) imgs_transition = lb.run_transition(list_nmb_branches, list_injection_strength, fixed_seeds=fixed_seeds) xxx #%% num_inference_steps = 30 # Number of diffusion interations list_nmb_branches = [2, 10, 50, 100, 200] # list_injection_strength = list(np.linspace(0.5, 0.95, 4)) # Branching structure: how deep is the blending list_injection_strength.insert(0, 0.0) width = 512 height = 512 guidance_scale = 5 fps = 30 duration_single_trans = 20 width = 512 height = 512 lb = LatentBlending(pipe, device, height, width, num_inference_steps, guidance_scale) list_prompts = [] list_prompts.append("surrealistic statue made of glitter and dirt, standing in a lake, atmospheric light, strange glow") list_prompts.append("statue of a mix between a tree and human, made of marble, incredibly detailed") list_prompts.append("weird statue of a frog monkey, many colors, standing next to the ruins of an ancient city") list_prompts.append("statue made of hot metal, bizzarre, dark clouds in the sky") list_prompts.append("statue of a spider that looked like a human") list_prompts.append("statue of a bird that looked like a scorpion") list_prompts.append("statue of an ancient cybernetic messenger annoucing good news, golden, futuristic") list_seeds = [234187386, 422209351, 241845736, 28652396, 783279867, 831049796, 234903931] fp_movie = "/home/lugo/tmp/latentblending/bubua.mp4" ms = MovieSaver(fp_movie, fps=fps) lb.run_multi_transition( list_prompts, list_seeds, list_nmb_branches, list_injection_strength=list_injection_strength, ms=ms, fps=fps, duration_single_trans=duration_single_trans ) #%% get good branching struct #%% #%% """ TODO Coding: RUNNING WITHOUT PROMPT! auto mode (quality settings) save value ranges, can it be trashed? set all variables in init! self.img2... TODO Other: github write text requirements make graphic explaining make colab license twitter et al """