jptv/LLM-grounded-diffusion

1version = "v3.0"2 3import torch4import numpy as np5import models6import utils7from models import pipelines, sam8from utils import parse, latents9from shared import model_dict, sam_model_dict, DEFAULT_SO_NEGATIVE_PROMPT, DEFAULT_OVERALL_NEGATIVE_PROMPT10import gc11 12verbose = False13 14vae, tokenizer, text_encoder, unet, dtype = model_dict.vae, model_dict.tokenizer, model_dict.text_encoder, model_dict.unet, model_dict.dtype15 16model_dict.update(sam_model_dict)17 18 19# Hyperparams20height = 512  # default height of Stable Diffusion21width = 512  # default width of Stable Diffusion22H, W = height // 8, width // 8 # size of the latent23guidance_scale = 7.5  # Scale for classifier-free guidance24 25# batch size that is not 1 is not supported26overall_batch_size = 127 28# discourage masks with confidence below29discourage_mask_below_confidence = 0.8530 31# discourage masks with iou (with coarse binarized attention mask) below32discourage_mask_below_coarse_iou = 0.2533 34run_ind = None35 36 37def generate_single_object_with_box_batch(prompts, bboxes, phrases, words, input_latents_list, input_embeddings, 38                                    sam_refine_kwargs, num_inference_steps, gligen_scheduled_sampling_beta=0.3, 39                                    verbose=False, scheduler_key=None, visualize=True, batch_size=None):40    # batch_size=None: does not limit the batch size (pass all input together)41    42    # prompts and words are not used since we don't have cross-attention control in this function43    44    input_latents = torch.cat(input_latents_list, dim=0)45    46    # We need to "unsqueeze" to tell that we have only one box and phrase in each batch item47    bboxes, phrases = [[item] for item in bboxes], [[item] for item in phrases]48    49    input_len = len(bboxes)50    assert len(bboxes) == len(phrases), f"{len(bboxes)} != {len(phrases)}"51    52    if batch_size is None:53        batch_size = input_len54    55    run_times = int(np.ceil(input_len / batch_size))56    mask_selected_list, single_object_pil_images_box_ann, latents_all = [], [], []57    for batch_idx in range(run_times):58        input_latents_batch, bboxes_batch, phrases_batch = input_latents[batch_idx * batch_size:(batch_idx + 1) * batch_size], \59            bboxes[batch_idx * batch_size:(batch_idx + 1) * batch_size], phrases[batch_idx * batch_size:(batch_idx + 1) * batch_size]60        input_embeddings_batch = input_embeddings[0], input_embeddings[1][batch_idx * batch_size:(batch_idx + 1) * batch_size]61        62        _, single_object_images_batch, single_object_pil_images_box_ann_batch, latents_all_batch = pipelines.generate_gligen(63            model_dict, input_latents_batch, input_embeddings_batch, num_inference_steps, bboxes_batch, phrases_batch, gligen_scheduled_sampling_beta=gligen_scheduled_sampling_beta, 64            guidance_scale=guidance_scale, return_saved_cross_attn=False,65            return_box_vis=True, save_all_latents=True, batched_condition=True, scheduler_key=scheduler_key66        )67        68        gc.collect()69        torch.cuda.empty_cache()70        71        # `sam_refine_boxes` also calls `empty_cache` so we don't need to explicitly empty the cache again.72        mask_selected, _ = sam.sam_refine_boxes(sam_input_images=single_object_images_batch, boxes=bboxes_batch, model_dict=model_dict, verbose=verbose, **sam_refine_kwargs)73        74        mask_selected_list.append(np.array(mask_selected)[:, 0])75        single_object_pil_images_box_ann.append(single_object_pil_images_box_ann_batch)76        latents_all.append(latents_all_batch)77    78    single_object_pil_images_box_ann, latents_all = sum(single_object_pil_images_box_ann, []), torch.cat(latents_all, dim=1)79    80    # mask_selected_list: List(batch)[List(image)[List(box)[Array of shape (64, 64)]]]81    82    mask_selected = np.concatenate(mask_selected_list, axis=0)83    mask_selected = mask_selected.reshape((-1, *mask_selected.shape[-2:]))84    85    assert mask_selected.shape[0] == input_latents.shape[0], f"{mask_selected.shape[0]} != {input_latents.shape[0]}"86    87    print(mask_selected.shape)88    89    mask_selected_tensor = torch.tensor(mask_selected)90    91    latents_all = latents_all.transpose(0,1)[:,:,None,...]92    93    gc.collect()94    torch.cuda.empty_cache()95    96    return latents_all, mask_selected_tensor, single_object_pil_images_box_ann97 98def get_masked_latents_all_list(so_prompt_phrase_word_box_list, input_latents_list, so_input_embeddings, verbose=False, **kwargs):99    latents_all_list, mask_tensor_list = [], []100   101    if not so_prompt_phrase_word_box_list:102        return latents_all_list, mask_tensor_list103    104    prompts, bboxes, phrases, words = [], [], [], []105 106    for prompt, phrase, word, box in so_prompt_phrase_word_box_list:107        prompts.append(prompt)108        bboxes.append(box)109        phrases.append(phrase)110        words.append(word)111    112    latents_all_list, mask_tensor_list, so_img_list = generate_single_object_with_box_batch(prompts, bboxes, phrases, words, input_latents_list, input_embeddings=so_input_embeddings, verbose=verbose, **kwargs)113 114    return latents_all_list, mask_tensor_list, so_img_list115 116 117# Note: need to keep the supervision, especially the box corrdinates, corresponds to each other in single object and overall.118 119def run(120    spec, bg_seed = 1, overall_prompt_override="", fg_seed_start = 20, frozen_step_ratio=0.4, gligen_scheduled_sampling_beta = 0.3, num_inference_steps = 20,121    so_center_box = False, fg_blending_ratio = 0.1, scheduler_key='dpm_scheduler', so_negative_prompt = DEFAULT_SO_NEGATIVE_PROMPT, overall_negative_prompt = DEFAULT_OVERALL_NEGATIVE_PROMPT, so_horizontal_center_only = True, 122    align_with_overall_bboxes = False, horizontal_shift_only = True, use_autocast = False, so_batch_size = None123):124    """    125    so_center_box: using centered box in single object generation126    so_horizontal_center_only: move to the center horizontally only127    128    align_with_overall_bboxes: Align the center of the mask, latents, and cross-attention with the center of the box in overall bboxes129    horizontal_shift_only: only shift horizontally for the alignment of mask, latents, and cross-attention130    """131    132    print("generation:", spec, bg_seed, fg_seed_start, frozen_step_ratio, gligen_scheduled_sampling_beta)133    134    frozen_step_ratio = min(max(frozen_step_ratio, 0.), 1.)135    frozen_steps = int(num_inference_steps * frozen_step_ratio)136 137    if True:138        so_prompt_phrase_word_box_list, overall_prompt, overall_phrases_words_bboxes = parse.convert_spec(spec, height, width, verbose=verbose)139 140    if overall_prompt_override and overall_prompt_override.strip():141        overall_prompt = overall_prompt_override.strip()142 143    overall_phrases, overall_words, overall_bboxes = [item[0] for item in overall_phrases_words_bboxes], [item[1] for item in overall_phrases_words_bboxes], [item[2] for item in overall_phrases_words_bboxes]144 145    # The so box is centered but the overall boxes are not (since we need to place to the right place).146    if so_center_box:147        so_prompt_phrase_word_box_list = [(prompt, phrase, word, utils.get_centered_box(bbox, horizontal_center_only=so_horizontal_center_only)) for prompt, phrase, word, bbox in so_prompt_phrase_word_box_list]148        if verbose:149            print(f"centered so_prompt_phrase_word_box_list: {so_prompt_phrase_word_box_list}")150    so_boxes = [item[-1] for item in so_prompt_phrase_word_box_list]151 152    sam_refine_kwargs = dict(153        discourage_mask_below_confidence=discourage_mask_below_confidence, discourage_mask_below_coarse_iou=discourage_mask_below_coarse_iou,154        height=height, width=width, H=H, W=W155    )156    157    # Note that so and overall use different negative prompts158 159    with torch.autocast("cuda", enabled=use_autocast):160        so_prompts = [item[0] for item in so_prompt_phrase_word_box_list]161        if so_prompts:162            so_input_embeddings = models.encode_prompts(prompts=so_prompts, tokenizer=tokenizer, text_encoder=text_encoder, negative_prompt=so_negative_prompt, one_uncond_input_only=True)163        else:164            so_input_embeddings = []165 166        overall_input_embeddings = models.encode_prompts(prompts=[overall_prompt], tokenizer=tokenizer, negative_prompt=overall_negative_prompt, text_encoder=text_encoder)167        168        input_latents_list, latents_bg = latents.get_input_latents_list(169            model_dict, bg_seed=bg_seed, fg_seed_start=fg_seed_start, 170            so_boxes=so_boxes, fg_blending_ratio=fg_blending_ratio, height=height, width=width, verbose=False171        )172        latents_all_list, mask_tensor_list, so_img_list = get_masked_latents_all_list(173            so_prompt_phrase_word_box_list, input_latents_list, 174            gligen_scheduled_sampling_beta=gligen_scheduled_sampling_beta,175            sam_refine_kwargs=sam_refine_kwargs, so_input_embeddings=so_input_embeddings, num_inference_steps=num_inference_steps, scheduler_key=scheduler_key, verbose=verbose, batch_size=so_batch_size176        )177 178        179 180        composed_latents, foreground_indices, offset_list = latents.compose_latents_with_alignment(181            model_dict, latents_all_list, mask_tensor_list, num_inference_steps, 182            overall_batch_size, height, width, latents_bg=latents_bg, 183            align_with_overall_bboxes=align_with_overall_bboxes, overall_bboxes=overall_bboxes,184            horizontal_shift_only=horizontal_shift_only185        )186        187        overall_bboxes_flattened, overall_phrases_flattened = [], []188        for overall_bboxes_item, overall_phrase in zip(overall_bboxes, overall_phrases):189            for overall_bbox in overall_bboxes_item:190                overall_bboxes_flattened.append(overall_bbox)191                overall_phrases_flattened.append(overall_phrase)192 193        # Generate with composed latents194 195        # Foreground should be frozen196        frozen_mask = foreground_indices != 0197        198        regen_latents, images = pipelines.generate_gligen(199            model_dict, composed_latents, overall_input_embeddings, num_inference_steps, 200            overall_bboxes_flattened, overall_phrases_flattened, guidance_scale=guidance_scale,201            gligen_scheduled_sampling_beta=gligen_scheduled_sampling_beta,202            frozen_steps=frozen_steps, frozen_mask=frozen_mask, scheduler_key=scheduler_key203        )204 205        print(f"Generation with spatial guidance from input latents and first {frozen_steps} steps frozen (directly from the composed latents input)")206        print("Generation from composed latents (with semantic guidance)")207 208        # display(Image.fromarray(images[0]), "img", run_ind)209        210    gc.collect()211    torch.cuda.empty_cache()212        213    return images[0], so_img_list214 215