import gradio as gr import spaces import torch import torchvision from torchvision import transforms import numpy as np import os from PIL import Image import rembg from huggingface_hub import hf_hub_download """ Generate a large batch of image samples from a model and save them as a large numpy array. This can be used to produce samples for FID evaluation. """ import argparse import json import sys import os sys.path.append('.') from pdb import set_trace as st import imageio import numpy as np import torch as th import torch.distributed as dist from guided_diffusion import dist_util, logger from guided_diffusion.script_util import ( NUM_CLASSES, model_and_diffusion_defaults, create_model_and_diffusion, add_dict_to_argparser, args_to_dict, continuous_diffusion_defaults, control_net_defaults, ) th.backends.cuda.matmul.allow_tf32 = True th.backends.cudnn.allow_tf32 = True th.backends.cudnn.enabled = True from pathlib import Path from tqdm import tqdm, trange import dnnlib from nsr.train_util_diffusion import TrainLoop3DDiffusion as TrainLoop from guided_diffusion.continuous_diffusion import make_diffusion as make_sde_diffusion import nsr import nsr.lsgm from nsr.script_util import create_3DAE_model, encoder_and_nsr_defaults, loss_defaults, AE_with_Diffusion, rendering_options_defaults, eg3d_options_default, dataset_defaults from datasets.shapenet import load_eval_data from torch.utils.data import Subset from datasets.eg3d_dataset import init_dataset_kwargs from transport.train_utils import parse_transport_args from utils.infer_utils import remove_background, resize_foreground SEED = 0 def resize_to_224(img): img = transforms.functional.resize(img, 224, interpolation=transforms.InterpolationMode.LANCZOS) return img def set_white_background(image): image = np.array(image).astype(np.float32) / 255.0 mask = image[:, :, 3:4] image = image[:, :, :3] * mask + (1 - mask) image = Image.fromarray((image * 255.0).astype(np.uint8)) return image def check_input_image(input_image): if input_image is None: raise gr.Error("No image uploaded!") def main(args): # args.rendering_kwargs = rendering_options_defaults(args) dist_util.setup_dist(args) logger.configure(dir=args.logdir) th.cuda.empty_cache() th.cuda.manual_seed_all(SEED) np.random.seed(SEED) # * set denoise model args logger.log("creating model and diffusion...") args.img_size = [args.image_size_encoder] # ! no longer required for LDM # args.denoise_in_channels = args.out_chans # args.denoise_out_channels = args.out_chans args.image_size = args.image_size_encoder # 224, follow the triplane size denoise_model, diffusion = create_model_and_diffusion( **args_to_dict(args, model_and_diffusion_defaults().keys())) # if 'cldm' in args.trainer_name: # assert isinstance(denoise_model, tuple) # denoise_model, controlNet = denoise_model # controlNet.to(dist_util.dev()) # controlNet.train() # else: # controlNet = None opts = eg3d_options_default() if args.sr_training: args.sr_kwargs = dnnlib.EasyDict( channel_base=opts.cbase, channel_max=opts.cmax, fused_modconv_default='inference_only', use_noise=True ) # ! close noise injection? since noise_mode='none' in eg3d # denoise_model.load_state_dict( # dist_util.load_state_dict(args.ddpm_model_path, map_location="cpu")) denoise_model.to(dist_util.dev()) if args.use_fp16: denoise_model.convert_to_fp16() denoise_model.eval() # * auto-encoder reconstruction model logger.log("creating 3DAE...") auto_encoder = create_3DAE_model( **args_to_dict(args, encoder_and_nsr_defaults().keys())) auto_encoder.to(dist_util.dev()) auto_encoder.eval() # TODO, how to set the scale? logger.log("create dataset") if args.objv_dataset: from datasets.g_buffer_objaverse import load_data, load_eval_data, load_memory_data, load_wds_data else: # shapenet from datasets.shapenet import load_data, load_eval_data, load_memory_data # load data if i23d if args.i23d: data = load_eval_data( file_path=args.eval_data_dir, batch_size=args.eval_batch_size, reso=args.image_size, reso_encoder=args.image_size_encoder, # 224 -> 128 num_workers=args.num_workers, load_depth=True, # for evaluation preprocess=auto_encoder.preprocess, **args_to_dict(args, dataset_defaults().keys())) else: data = None # t23d sampling, only caption required TrainLoop = { 'sgm_legacy': nsr.lsgm.sgm_DiffusionEngine.DiffusionEngineLSGM, 'flow_matching': nsr.lsgm.flow_matching_trainer.FlowMatchingEngine, }[args.trainer_name] # continuous sde_diffusion = None auto_encoder.decoder.rendering_kwargs = args.rendering_kwargs training_loop_class = TrainLoop(rec_model=auto_encoder, denoise_model=denoise_model, control_model=None, # to remove diffusion=diffusion, sde_diffusion=sde_diffusion, loss_class=None, data=data, eval_data=None, **vars(args)) @spaces.GPU() def reconstruct_and_export(*args, **kwargs): return training_loop_class.eval_i23d_and_export(*args, **kwargs) css = """ h1 { text-align: center; display:block; } """ def preprocess(input_image, preprocess_background=True, foreground_ratio=0.85): if preprocess_background: rembg_session = rembg.new_session() image = input_image.convert("RGB") image = remove_background(image, rembg_session) image = resize_foreground(image, foreground_ratio) image = set_white_background(image) else: image = input_image if image.mode == "RGBA": image = set_white_background(image) image = resize_to_224(image) return image with gr.Blocks(css=css) as demo: gr.Markdown( """ # LN3Diff (Scalable Latent Neural Fields Diffusion for Speedy 3D Generation) **LN3Diff (ECCV 2024)** [[code](https://github.com/NIRVANALAN/LN3Diff), [project page](https://nirvanalan.github.io/projects/ln3diff/)] is a scalable 3D latent diffusion model that supports speedy 3D assets generation. It first trains a 3D VAE on **Objaverse**, which compress each 3D asset into a compact 3D-aware latent. After that, a image/text-conditioned diffusion model is trained following LDM paradigm. The model used in the demo adopts DiT-L/2 architecture and flow-matching framework, and supports single-image condition. It is trained on 8 A100 GPUs for 1M iterations with batch size 256. Locally, on an NVIDIA A100/A10 GPU, each image-conditioned diffusion generation can be done in 10~20 seconds (time varies due to the adaptive-step ODE solver used in flow-mathcing.) Upload an image of an object or click on one of the provided examples to see how the LN3Diff works. The 3D viewer will render a .obj object exported from the triplane, where the mesh resolution and iso-surface can be set manually. For best results run the demo locally and render locally - to do so, clone the [main repository](https://github.com/NIRVANALAN/LN3Diff). """ ) with gr.Row(variant="panel"): with gr.Column(): with gr.Row(): input_image = gr.Image( label="Input Image", image_mode="RGBA", sources="upload", type="pil", elem_id="content_image", ) processed_image = gr.Image(label="Processed Image", interactive=False) # params with gr.Row(): with gr.Column(): with gr.Row(): # with gr.Group(): unconditional_guidance_scale = gr.Number( label="CFG-scale", value=4.0, interactive=True, ) seed = gr.Number( label="Seed", value=42, interactive=True, ) num_steps = gr.Number( label="ODE Sampling Steps", value=250, interactive=True, ) # with gr.Column(): with gr.Row(): mesh_size = gr.Number( label="Mesh Resolution", value=192, interactive=True, ) mesh_thres = gr.Number( label="Mesh Iso-surface", value=10, interactive=True, ) with gr.Row(): with gr.Group(): preprocess_background = gr.Checkbox( label="Remove Background", value=True ) with gr.Row(): submit = gr.Button("Generate", elem_id="generate", variant="primary") with gr.Row(variant="panel"): gr.Examples( examples=[ str(path) for path in sorted(Path('./assets/i23d_examples').glob('**/*.png')) ], inputs=[input_image], cache_examples=False, label="Examples", examples_per_page=20, ) with gr.Column(): with gr.Row(): with gr.Tab("Reconstruction"): with gr.Column(): output_video = gr.Video(value=None, width=384, label="Rendered Video", autoplay=True, loop=True) output_model = gr.Model3D( height=384, clear_color=(1,1,1,1), label="Output Model", interactive=False ) gr.Markdown( """ ## Comments: 1. The sampling time varies since ODE-based sampling method (dopri5 by default) has adaptive internal step, and reducing sampling steps may not reduce the overal sampling time. Sampling steps=250 is the emperical value that works well in most cases. 2. The 3D viewer shows a colored .glb mesh extracted from volumetric tri-plane, and may differ slightly with the volume rendering result. 3. If you find your result unsatisfying, tune the CFG scale and change the random seed. Usually slightly increase the CFG value can lead to better performance. 3. Known limitations include: - Texture details missing: since our VAE is trained on 192x192 resolution due the the resource constraints, the texture details generated by the final 3D-LDM may be blurry. We will keep improving the performance in the future. 4. Regarding reconstruction performance, our model is slightly inferior to state-of-the-art multi-view LRM-based method (e.g. InstantMesh), but offers much better diversity, flexibility and editing potential due to the intrinsic nature of diffusion model. ## How does it work? LN3Diff is a feedforward 3D Latent Diffusion Model that supports direct 3D asset generation via diffusion sampling. Compared to SDS-based ([DreamFusion](https://dreamfusion3d.github.io/)), mulit-view generation-based ([MVDream](https://arxiv.org/abs/2308.16512), [Zero123++](https://github.com/SUDO-AI-3D/zero123plus), [Instant3D](https://instant-3d.github.io/)) and feedforward 3D reconstruction-based ([LRM](https://yiconghong.me/LRM/), [InstantMesh](https://github.com/TencentARC/InstantMesh), [LGM](https://github.com/3DTopia/LGM)), LN3Diff supports feedforward 3D generation with a unified framework. Like 2D/Video AIGC pipeline, LN3Diff first trains a 3D-VAE and then conduct LDM training (text/image conditioned) on the learned latent space. Some related methods from the industry ([Shape-E](https://github.com/openai/shap-e), [CLAY](https://github.com/CLAY-3D/OpenCLAY), [Meta 3D Gen](https://arxiv.org/abs/2303.05371)) also follow the same paradigm. Though currently the performance of the origin 3D LDM's works are overall inferior to reconstruction-based methods, we believe the proposed method has much potential and scales better with more data and compute resources, and may yield better 3D editing performance due to its compatability with diffusion model. For more results see the [project page](https://szymanowiczs.github.io/splatter-image) and the [ECCV article](https://arxiv.org/pdf/2403.12019). """ ) submit.click(fn=check_input_image, inputs=[input_image]).success( fn=preprocess, inputs=[input_image, preprocess_background], outputs=[processed_image], ).success( # fn=reconstruct_and_export, # inputs=[processed_image], # outputs=[output_model, output_video], fn=reconstruct_and_export, inputs=[processed_image, num_steps, seed, mesh_size, mesh_thres, unconditional_guidance_scale], outputs=[output_video, output_model], ) demo.queue(max_size=1) demo.launch(share=True) # training_loop_class.eval_i23d_and_export( # # prompt=args.prompt, # # prompt=prompt, # unconditional_guidance_scale=args. # unconditional_guidance_scale, # # unconditional_guidance_scale=unconditional_guidance_scale, # # use_ddim=args.use_ddim, # # save_img=args.save_img, # # use_train_trajectory=args.use_train_trajectory, # camera=camera, # num_instances=args.num_instances, # num_samples=args.num_samples, # export_mesh=True, # idx_to_render=seeds, # ) def create_argparser(): defaults = dict( image_size_encoder=224, triplane_scaling_divider=1.0, # divide by this value diffusion_input_size=-1, trainer_name='adm', use_amp=False, # triplane_scaling_divider=1.0, # divide by this value # * sampling flags clip_denoised=False, num_samples=10, num_instances=10, # for i23d, loop different condition use_ddim=False, ddpm_model_path="", cldm_model_path="", rec_model_path="", # * eval logging flags logdir="/mnt/lustre/yslan/logs/nips23/", data_dir="", eval_data_dir="", eval_batch_size=1, num_workers=1, # * training flags for loading TrainingLoop class overfitting=False, image_size=128, iterations=150000, schedule_sampler="uniform", anneal_lr=False, lr=5e-5, weight_decay=0.0, lr_anneal_steps=0, batch_size=1, microbatch=-1, # -1 disables microbatches ema_rate="0.9999", # comma-separated list of EMA values log_interval=50, eval_interval=2500, save_interval=10000, resume_checkpoint="", resume_cldm_checkpoint="", resume_checkpoint_EG3D="", use_fp16=False, fp16_scale_growth=1e-3, load_submodule_name='', # for loading pretrained auto_encoder model ignore_resume_opt=False, freeze_ae=False, denoised_ae=True, # inference prompt prompt="a red chair", interval=1, save_img=False, use_train_trajectory= False, # use train trajectory to sample images for fid calculation unconditional_guidance_scale=1.0, use_eos_feature=False, export_mesh=False, cond_key='caption', allow_tf32=True, ) defaults.update(model_and_diffusion_defaults()) defaults.update(encoder_and_nsr_defaults()) # type: ignore defaults.update(loss_defaults()) defaults.update(continuous_diffusion_defaults()) defaults.update(control_net_defaults()) defaults.update(dataset_defaults()) parser = argparse.ArgumentParser() add_dict_to_argparser(parser, defaults) parse_transport_args(parser) return parser if __name__ == "__main__": # os.environ["TORCH_CPP_LOG_LEVEL"] = "INFO" # os.environ["NCCL_DEBUG"] = "INFO" os.environ[ "TORCH_DISTRIBUTED_DEBUG"] = "DETAIL" # set to DETAIL for runtime logging. args = create_argparser().parse_args() args.local_rank = int(os.environ["LOCAL_RANK"]) args.gpus = th.cuda.device_count() args.rendering_kwargs = rendering_options_defaults(args) # main(args) zero = torch.Tensor([0]).cuda() print(zero.device) # <-- 'cpu' 🤔 @spaces.GPU def greet(n): print(zero.device) # <-- 'cuda:0' 🤗 return f"Hello {zero + n} Tensor" demo = gr.Interface(fn=greet, inputs=gr.Number(), outputs=gr.Text())