import os from typing import List, Tuple PWD = os.path.dirname(__file__) import subprocess subprocess.run("pip install flash-attn --no-build-isolation", env={"FLASH_ATTENTION_SKIP_CUDA_BUILD": "TRUE"}, shell=True) try: import os from huggingface_hub import login # Try to login with token from environment variable hf_token = os.environ["HF_TOKEN"] if hf_token: login(token=hf_token) print("✅ Authenticated with Hugging Face") else: print("No HF_TOKEN found, trying without authentication...") except Exception as e: print(f"Authentication failed: {e}") # download checkpoints from download_checkpoints import main as download_checkpoints os.makedirs("./checkpoints", exist_ok=True) download_checkpoints(hf_token="", output_dir="./checkpoints", model="7b_av") os.environ["TOKENIZERS_PARALLELISM"] = "false" # Workaround to suppress MP warning import copy import json import random from io import BytesIO import gradio as gr import torch from cosmos_transfer1.checkpoints import ( BASE_7B_CHECKPOINT_AV_SAMPLE_PATH, BASE_7B_CHECKPOINT_PATH, EDGE2WORLD_CONTROLNET_DISTILLED_CHECKPOINT_PATH, ) from cosmos_transfer1.diffusion.inference.inference_utils import ( validate_controlnet_specs, ) from cosmos_transfer1.diffusion.inference.preprocessors import Preprocessors from cosmos_transfer1.diffusion.inference.world_generation_pipeline import ( DiffusionControl2WorldGenerationPipeline, DistilledControl2WorldGenerationPipeline, ) from cosmos_transfer1.utils import log, misc from cosmos_transfer1.utils.io import read_prompts_from_file, save_video from helper import parse_arguments torch.enable_grad(False) torch.serialization.add_safe_globals([BytesIO]) def inference(cfg, control_inputs) -> Tuple[List[str], List[str]]: video_paths = [] prompt_paths = [] control_inputs = validate_controlnet_specs(cfg, control_inputs) misc.set_random_seed(cfg.seed) device_rank = 0 process_group = None if cfg.num_gpus > 1: from cosmos_transfer1.utils import distributed from megatron.core import parallel_state distributed.init() parallel_state.initialize_model_parallel(context_parallel_size=cfg.num_gpus) process_group = parallel_state.get_context_parallel_group() device_rank = distributed.get_rank(process_group) preprocessors = Preprocessors() if cfg.use_distilled: assert not cfg.is_av_sample checkpoint = EDGE2WORLD_CONTROLNET_DISTILLED_CHECKPOINT_PATH pipeline = DistilledControl2WorldGenerationPipeline( checkpoint_dir=cfg.checkpoint_dir, checkpoint_name=checkpoint, offload_network=cfg.offload_diffusion_transformer, offload_text_encoder_model=cfg.offload_text_encoder_model, offload_guardrail_models=cfg.offload_guardrail_models, guidance=cfg.guidance, num_steps=cfg.num_steps, fps=cfg.fps, seed=cfg.seed, num_input_frames=cfg.num_input_frames, control_inputs=control_inputs, sigma_max=cfg.sigma_max, blur_strength=cfg.blur_strength, canny_threshold=cfg.canny_threshold, upsample_prompt=cfg.upsample_prompt, offload_prompt_upsampler=cfg.offload_prompt_upsampler, process_group=process_group, ) else: checkpoint = BASE_7B_CHECKPOINT_AV_SAMPLE_PATH if cfg.is_av_sample else BASE_7B_CHECKPOINT_PATH # Initialize transfer generation model pipeline pipeline = DiffusionControl2WorldGenerationPipeline( checkpoint_dir=cfg.checkpoint_dir, checkpoint_name=checkpoint, offload_network=cfg.offload_diffusion_transformer, offload_text_encoder_model=cfg.offload_text_encoder_model, offload_guardrail_models=cfg.offload_guardrail_models, guidance=cfg.guidance, num_steps=cfg.num_steps, fps=cfg.fps, seed=cfg.seed, num_input_frames=cfg.num_input_frames, control_inputs=control_inputs, sigma_max=cfg.sigma_max, blur_strength=cfg.blur_strength, canny_threshold=cfg.canny_threshold, upsample_prompt=cfg.upsample_prompt, offload_prompt_upsampler=cfg.offload_prompt_upsampler, process_group=process_group, ) if cfg.batch_input_path: log.info(f"Reading batch inputs from path: {cfg.batch_input_path}") prompts = read_prompts_from_file(cfg.batch_input_path) else: # Single prompt case prompts = [{"prompt": cfg.prompt, "visual_input": cfg.input_video_path}] batch_size = cfg.batch_size if hasattr(cfg, "batch_size") else 1 if any("upscale" in control_input for control_input in control_inputs) and batch_size > 1: batch_size = 1 log.info("Setting batch_size=1 as upscale does not support batch generation") os.makedirs(cfg.video_save_folder, exist_ok=True) for batch_start in range(0, len(prompts), batch_size): # Get current batch batch_prompts = prompts[batch_start : batch_start + batch_size] actual_batch_size = len(batch_prompts) # Extract batch data batch_prompt_texts = [p.get("prompt", None) for p in batch_prompts] batch_video_paths = [p.get("visual_input", None) for p in batch_prompts] batch_control_inputs = [] for i, input_dict in enumerate(batch_prompts): current_prompt = input_dict.get("prompt", None) current_video_path = input_dict.get("visual_input", None) if cfg.batch_input_path: video_save_subfolder = os.path.join(cfg.video_save_folder, f"video_{batch_start+i}") os.makedirs(video_save_subfolder, exist_ok=True) else: video_save_subfolder = cfg.video_save_folder current_control_inputs = copy.deepcopy(control_inputs) if "control_overrides" in input_dict: for hint_key, override in input_dict["control_overrides"].items(): if hint_key in current_control_inputs: current_control_inputs[hint_key].update(override) else: log.warning(f"Ignoring unknown control key in override: {hint_key}") # if control inputs are not provided, run respective preprocessor (for seg and depth) log.info("running preprocessor") preprocessors( current_video_path, current_prompt, current_control_inputs, video_save_subfolder, cfg.regional_prompts if hasattr(cfg, "regional_prompts") else None, ) batch_control_inputs.append(current_control_inputs) regional_prompts = [] region_definitions = [] if hasattr(cfg, "regional_prompts") and cfg.regional_prompts: log.info(f"regional_prompts: {cfg.regional_prompts}") for regional_prompt in cfg.regional_prompts: regional_prompts.append(regional_prompt["prompt"]) if "region_definitions_path" in regional_prompt: log.info(f"region_definitions_path: {regional_prompt['region_definitions_path']}") region_definition_path = regional_prompt["region_definitions_path"] if isinstance(region_definition_path, str) and region_definition_path.endswith(".json"): with open(region_definition_path, "r") as f: region_definitions_json = json.load(f) region_definitions.extend(region_definitions_json) else: region_definitions.append(region_definition_path) if hasattr(pipeline, "regional_prompts"): pipeline.regional_prompts = regional_prompts if hasattr(pipeline, "region_definitions"): pipeline.region_definitions = region_definitions # Generate videos in batch batch_outputs = pipeline.generate( prompt=batch_prompt_texts, video_path=batch_video_paths, negative_prompt=cfg.negative_prompt, control_inputs=batch_control_inputs, save_folder=video_save_subfolder, batch_size=actual_batch_size, ) if batch_outputs is None: log.critical("Guardrail blocked generation for entire batch.") continue videos, final_prompts = batch_outputs for i, (video, prompt) in enumerate(zip(videos, final_prompts)): if cfg.batch_input_path: video_save_subfolder = os.path.join(cfg.video_save_folder, f"video_{batch_start+i}") video_save_path = os.path.join(video_save_subfolder, "output.mp4") prompt_save_path = os.path.join(video_save_subfolder, "prompt.txt") else: video_save_path = os.path.join(cfg.video_save_folder, f"{cfg.video_save_name}.mp4") prompt_save_path = os.path.join(cfg.video_save_folder, f"{cfg.video_save_name}.txt") # Save video and prompt if device_rank == 0: os.makedirs(os.path.dirname(video_save_path), exist_ok=True) save_video( video=video, fps=cfg.fps, H=video.shape[1], W=video.shape[2], video_save_quality=5, video_save_path=video_save_path, ) video_paths.append(video_save_path) # Save prompt to text file alongside video with open(prompt_save_path, "wb") as f: f.write(prompt.encode("utf-8")) prompt_paths.append(prompt_save_path) log.info(f"Saved video to {video_save_path}") log.info(f"Saved prompt to {prompt_save_path}") # clean up properly if cfg.num_gpus > 1: parallel_state.destroy_model_parallel() import torch.distributed as dist dist.destroy_process_group() return video_paths, prompt_paths def generate_video( hdmap_video_input, lidar_video_input, prompt, negative_prompt="The video captures a series of frames showing ugly scenes, static with no motion, motion blur, over-saturation, shaky footage, low resolution, grainy texture, pixelated images, poorly lit areas, underexposed and overexposed scenes, poor color balance, washed out colors, choppy sequences, jerky movements, low frame rate, artifacting, color banding, unnatural transitions, outdated special effects, fake elements, unconvincing visuals, poorly edited content, jump cuts, visual noise, and flickering. Overall, the video is of poor quality.", # noqa: E501 seed=42, randomize_seed=False, progress=gr.Progress(track_tqdm=True), ): if randomize_seed: actual_seed = random.randint(0, 1000000) else: actual_seed = seed args, control_inputs = parse_arguments( controlnet_specs_in={ "hdmap": {"control_weight": 0.3, "input_control": hdmap_video_input}, "lidar": {"control_weight": 0.7, "input_control": lidar_video_input}, }, checkpoint_dir="./cosmos-transfer1/checkpoints", prompt=prompt, negative_prompt=negative_prompt, sigma_max=80, offload_text_encoder_model=True, is_av_sample=True, num_gpus=1, seed=seed, ) videos, prompts = inference(args, control_inputs) video = videos[0] return video, video, actual_seed # Define the Gradio Blocks interface with gr.Blocks() as demo: gr.Markdown( """ # Cosmos-Transfer1-7B-Sample-AV """ ) with gr.Row(): with gr.Column(): hdmap_input = gr.Video(label="Input HD Map Video", format="mp4") lidar_input = gr.Video(label="Input LiDAR Video", format="mp4") prompt_input = gr.Textbox( label="Prompt", lines=5, value="A close-up shot captures a vibrant yellow scrubber vigorously working on a grimy plate, its bristles moving in circular motions to lift stubborn grease and food residue. The dish, once covered in remnants of a hearty meal, gradually reveals its original glossy surface. Suds form and bubble around the scrubber, creating a satisfying visual of cleanliness in progress. The sound of scrubbing fills the air, accompanied by the gentle clinking of the dish against the sink. As the scrubber continues its task, the dish transforms, gleaming under the bright kitchen lights, symbolizing the triumph of cleanliness over mess.", # noqa: E501 placeholder="Enter your descriptive prompt here...", ) negative_prompt_input = gr.Textbox( label="Negative Prompt", lines=3, value="The video captures a series of frames showing ugly scenes, static with no motion, motion blur, over-saturation, shaky footage, low resolution, grainy texture, pixelated images, poorly lit areas, underexposed and overexposed scenes, poor color balance, washed out colors, choppy sequences, jerky movements, low frame rate, artifacting, color banding, unnatural transitions, outdated special effects, fake elements, unconvincing visuals, poorly edited content, jump cuts, visual noise, and flickering. Overall, the video is of poor quality.", # noqa: E501 placeholder="Enter what you DON'T want to see in the image...", ) with gr.Row(): randomize_seed_checkbox = gr.Checkbox(label="Randomize Seed", value=True) seed_input = gr.Slider(minimum=0, maximum=1000000, value=1, step=1, label="Seed") generate_button = gr.Button("Generate Image") with gr.Column(): output_video = gr.Video(label="Generated Video", format="mp4") output_file = gr.File(label="Download Video") generate_button.click( fn=generate_video, inputs=[hdmap_input, lidar_input, prompt_input, negative_prompt_input, seed_input, randomize_seed_checkbox], outputs=[output_video, output_file, seed_input], ) if __name__ == "__main__": demo.launch()