init

README.md

@@ -1,14 +1,122 @@
----
-title: GaussianAnything AIGC3D
-emoji: 📈
-colorFrom: indigo
-colorTo: blue
-sdk: gradio
-sdk_version: 5.6.0
-app_file: app.py
-pinned: false
-license: other
-short_description: GaussianAnything generates high-quality and editable 2DGS.
----
-
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# GaussianAnything: arXiv 2024
+
+## setup the environment (the same env as LN3Diff)
+
+```bash
+conda create -n ga python=3.10
+conda activate ga
+pip intall -r requrements.txt # will install the surfel Gaussians environments automatically.
+```
+
+Then, install pytorch3d with 
+```bash
+pip install git+https://github.com/facebookresearch/pytorch3d.git@stable
+```
+
+
+### :dromedary_camel: TODO
+
+- [x] Release inference code and checkpoints.
+- [x] Release Training code.
+- [x] Release pre-extracted latent codes for 3D diffusion training.
+- [ ] Release Gradio Demo.
+- [ ] Release the evaluation code.
+- [ ] Lint the code.
+
+
+# Inference
+
+Be aware to change the $logdir in the bash file accordingly.
+
+To load the checkpoint automatically: please replace ```/mnt/sfs-common/yslan/open-source``` with ```yslan/GaussianAnything/ckpts/checkpoints```.
+
+
+
+## Text-2-3D:
+
+Please update the caption for 3D generation in ```datasets/caption-forpaper.txt```. T o change the number of samples to be generated, please change ```$num_samples``` in the bash file.
+
+**stage-1**:
+```
+bash shell_scripts/release/inference/t23d/stage1-t23d.sh
+```
+then, set the ```$stage_1_output_dir``` to the ```$logdir``` of the above stage.
+
+**stage-2**: 
+```
+bash shell_scripts/release/inference/t23d/stage2-t23d.sh
+```
+
+The results will be dumped to ```./logs/t23d/stage-2```
+
+## I23D (requires two stage generation):
+
+set the $data_dir accordingly. For some demo image, please download from [huggingfac.co/yslan/GaussianAnything/demo-img](https://huggingface.co/yslan/GaussianAnything/tree/main/demo-img).
+
+**stage-1**:
+```
+bash shell_scripts/release/inference/i23d/i23d-stage1.sh
+```
+
+then, set the $stage_1_output_dir to the $logdir of the above stage.
+
+**stage-2**: 
+```
+bash shell_scripts/release/inference/i23d/i23d-stage1.sh
+```
+
+## 3D VAE Reconstruction:
+
+To encode a 3D asset into the latent point cloud, please download the pre-trained VAE checkpoint from [huggingfac.co/yslan/gaussiananything/ckpts/vae/model_rec1965000.pt](https://huggingface.co/yslan/GaussianAnything/blob/main/ckpts/vae/model_rec1965000.pt) to ```./checkpoint/model_rec1965000.pt```.
+
+Then, run the inference script
+
+```bash 
+bash shell_scripts/release/inference/vae-3d.sh
+```
+
+This will encode the mulit-view 3D renderings in ```./assets/demo-image-for-i23d/for-vae-reconstruction/Animals/0``` into the point-cloud structured latent code, and export them (along with the 2dgs mesh) in ```./logs/latent_dir/```. The exported latent code will be used for efficient 3D diffusion training.
+
+
+
+# Training (Flow Matching 3D Generation)
+All the training is conducted on 8 A100 (80GiB) with BF16 enabled. For training on V100, please use FP32 training by setting ```--use_amp``` False in the bash file. Feel free to tune the ```$batch_size``` in the bash file accordingly to match your VRAM.
+
+To facilitate reproducing the performance, we have uploaded the pre-extracted poind cloud-structured latent codes to the [huggingfac.co/yslan/gaussiananything/dataset/latent.tar.gz](https://huggingface.co/yslan/GaussianAnything/blob/main/dataset/latent.tar.gz) (34GiB required). Please download the pre extracted point cloud latent codes, unzip and set the ```$mv_latent_dir``` in the bash file accordingly.
+
+
+## Text to 3D:
+Please donwload the 3D caption from hugging face [huggingfac.co/yslan/GaussianAnything/dataset/text_captions_3dtopia.json](https://huggingface.co/yslan/GaussianAnything/blob/main/dataset/text_captions_3dtopia.json), and put it under ```dataset```.
+
+
+Note that if you want to train a specific class of Objaverse, just manually change the code at ```datasets/g_buffer_objaverse.py:3043```.
+
+**stage-1 training (point cloud generation)**:
+
+```
+bash shell_scripts/release/train/stage2-t23d/t23d-pcd-gen.sh
+```
+
+**stage-2 training (point cloud-conditioned KL feature generation)**:
+
+```
+bash shell_scripts/release/train/stage2-t23d/t23d-klfeat-gen.sh
+```
+
+## (single-view) Image to 3D
+Please download g-buffer dataset first.
+
+**stage-1 training (point cloud generation)**:
+
+```
+bash shell_scripts/release/train/stage2-i23d/i23d-pcd-gen.sh
+```
+
+**stage-2 training (point cloud-conditioned KL feature generation)**:
+
+```
+bash shell_scripts/release/train/stage2-i23d/i23d-klfeat-gen.sh
+```
+
+<!-- # Training (3D-aware VAE)
+Since the  -->

app.py

@@ -0,0 +1,397 @@
+import argparse
+import spaces
+import json
+import sys
+sys.path.append('.')
+import torch
+import torchvision
+from torchvision import transforms
+import numpy as np
+
+import os
+import gc
+import dnnlib
+from omegaconf import OmegaConf
+from PIL import Image 
+from dnnlib.util import EasyDict
+
+import gradio as gr
+
+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 os
+
+
+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, 518, # required by dino.
+        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_1, args_2):
+
+    os.environ['MASTER_ADDR'] = 'localhost'
+    os.environ['MASTER_PORT'] = '12355'
+    os.environ["CUDA_VISIBLE_DEVICES"] = "0"
+    os.environ["RANK"] = "0"
+    os.environ["WORLD_SIZE"] = "1"
+
+    # args.rendering_kwargs = rendering_options_defaults(args)
+
+    dist_util.setup_dist(args_1)
+    logger.configure(dir=args_1.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_1.img_size = [args_1.image_size_encoder]
+    args_1.image_size = args_1.image_size_encoder  # 224, follow the triplane size
+
+    args_2.img_size = [args_2.image_size_encoder]
+    args_2.image_size = args_2.image_size_encoder  # 224, follow the triplane size
+
+    denoise_model_stage1, diffusion = create_model_and_diffusion(
+        **args_to_dict(args_1,
+                       model_and_diffusion_defaults().keys()))
+
+    denoise_model_stage2, diffusion = create_model_and_diffusion(
+        **args_to_dict(args_2,
+                       model_and_diffusion_defaults().keys()))
+
+    opts = eg3d_options_default()
+
+    denoise_model_stage1.to(dist_util.dev())
+    denoise_model_stage1.eval()
+    denoise_model_stage2.to(dist_util.dev())
+    denoise_model_stage2.eval()
+
+    # * auto-encoder reconstruction model
+    logger.log("creating 3DAE...")
+    auto_encoder = create_3DAE_model(
+        **args_to_dict(args_1,
+                       encoder_and_nsr_defaults().keys()))
+
+    auto_encoder.to(dist_util.dev())
+    auto_encoder.eval()
+
+    # faster inference
+    # denoise_model = denoise_model.to(th.bfloat16)
+    # auto_encoder = auto_encoder.to(th.bfloat16)
+
+    # TODO, how to set the scale?
+    logger.log("create dataset")
+
+    if args_1.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 = {
+        'flow_matching':
+        nsr.lsgm.flow_matching_trainer.FlowMatchingEngine,
+        'flow_matching_gs':  
+        nsr.lsgm.flow_matching_trainer.FlowMatchingEngine_gs, # slightly modified sampling and rendering for gs
+    }[args_1.trainer_name]
+
+    # continuous
+    sde_diffusion = None
+
+    auto_encoder.decoder.rendering_kwargs = args_1.rendering_kwargs
+    # stage_1_output_dir = args_2.stage_1_output_dir
+
+    training_loop_class_stage1 = TrainLoop(rec_model=auto_encoder,
+                                    denoise_model=denoise_model_stage1,
+                                    control_model=None, # to remove
+                                    diffusion=diffusion,
+                                    sde_diffusion=sde_diffusion,
+                                    loss_class=None,
+                                    data=data,
+                                    eval_data=None,
+                                    **args_1)
+
+    training_loop_class_stage2 = TrainLoop(rec_model=auto_encoder,
+                                    denoise_model=denoise_model_stage2,
+                                    control_model=None, # to remove
+                                    diffusion=diffusion,
+                                    sde_diffusion=sde_diffusion,
+                                    loss_class=None,
+                                    data=data,
+                                    eval_data=None,
+                                    **args_2)
+
+
+    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
+
+
+    @spaces.GPU(duration=50)
+    def cascaded_generation(processed_image, seed, cfg_scale):
+        # gc.collect()
+        # stage-1, generate pcd
+        stage_1_pcd = training_loop_class_stage1.eval_i23d_and_export_gradio(processed_image, seed, cfg_scale)
+        # stage-2, generate surfel Gaussians, tsdf mesh etc.
+        video_path, rgb_xyz_path, post_mesh_path = training_loop_class_stage2.eval_i23d_and_export_gradio(processed_image, seed, cfg_scale)
+        return video_path, rgb_xyz_path, post_mesh_path, stage_1_pcd
+
+    with gr.Blocks(css=css) as demo:
+        gr.Markdown(
+            """
+            # GaussianAnything: Interactive Point Cloud Latent Diffusion for 3D Generation
+            **GaussianAnything (arXiv 2024)** [[code](https://github.com/NIRVANALAN/GaussianAnything), [project page](https://nirvanalan.github.io/projects/GA/)] is a native 3D diffusion model that supports high-quality 2D Gaussians generation. 
+            It first trains a 3D VAE on **Objaverse**, which compress each 3D asset into a compact point cloud-structured latent. 
+            After that, a image/text-conditioned diffusion model is trained following LDM paradigm.
+            The model used in the demo adopts 3D DiT 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 within 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 GaussianAnything works.
+
+            The 3D viewer will render a .glb point cloud exported from the centers of the surfel Gaussians, and an integrated TSDF mesh.
+            For best results run the demo locally and render locally - to do so, clone the [main repository](https://github.com/NIRVANALAN/GaussianAnything).
+            """
+            )
+        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():
+
+                            cfg_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=False
+                        )
+                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/demo-image-for-i23d/instantmesh').glob('**/*.png'))
+                        ] + [str(path) for path in sorted(Path('./assets/demo-image-for-i23d/gso').glob('**/*.png'))],
+                        inputs=[input_image],
+                        cache_examples=False,
+                        label="Examples",
+                        examples_per_page=20,
+                    )
+
+            with gr.Column():
+                with gr.Row():
+                    with gr.Tab("Stage-2 Output"):
+                        with gr.Column():
+                            output_video = gr.Video(value=None, width=512, label="Rendered Video (2 LoDs)", autoplay=True, loop=True)
+                            # output_video = gr.Video(value=None, width=256, label="Rendered Video", autoplay=True)
+                            output_gs = gr.Model3D(
+                                height=256,
+                                label="2DGS Center",
+                                pan_speed=0.5,
+                                clear_color=(1,1,1,1), # loading glb file only.
+                            )
+                            output_model = gr.Model3D(
+                                height=256,
+                                label="TSDF Mesh",
+                                pan_speed=0.5,
+                                clear_color=(1,1,1,1), # loading tsdf ply files.
+                            )
+
+                    with gr.Tab("Stage-1 Output"):
+                        with gr.Column():
+                            output_model_stage1 = gr.Model3D(
+                                height=256,
+                                label="Stage-1",
+                                pan_speed=0.5,
+                                clear_color=(1,1,1,1), # loading tsdf ply files.
+                            )
+
+
+
+        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?
+
+            GaussianAnything is a native 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)), 
+            GaussianAnything supports feedforward 3D generation with a unified framework.
+            Like 2D/Video AIGC pipeline, GaussianAnything 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://nirvanalan.github.io/projects/GA/).
+            """
+        )
+
+        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=cascaded_generation,
+            inputs=[processed_image, seed, cfg_scale],
+            # inputs=[processed_image, num_steps, seed, mesh_size, mesh_thres, unconditional_guidance_scale, args.stage_1_output_dir],
+            outputs=[output_video, output_gs, output_model, output_model_stage1],
+        )
+
+    demo.queue(max_size=1)
+    demo.launch(share=True)
+
+if __name__ == "__main__":
+
+    os.environ[
+        "TORCH_DISTRIBUTED_DEBUG"] = "DETAIL"  # set to DETAIL for runtime logging.
+
+    with open('configs/gradio_i23d_stage2_args.json') as f:
+        args_2 = json.load(f)
+        args_2 = EasyDict(args_2)
+        args_2.local_rank = 0
+        args_2.gpus = 1
+
+    with open('configs/gradio_i23d_stage1_args.json') as f:
+        args_1 = json.load(f)
+        args_1 = EasyDict(args_1)
+        args_1.local_rank = 0
+        args_1.gpus = 1
+
+    main(args_1, args_2)

assets/demo-image-for-i23d/for-vae-reconstruction/Animals/0/10017/1/bbox.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b80359fdef16e0762ec4ec96046d23e4db42a6d9111a901fab4d63aa6327864c
+size 192

assets/demo-image-for-i23d/for-vae-reconstruction/Animals/0/10017/1/c.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:98b99b4d3319a30495c8726d4e9a32347938aa99e918fa86401da3eeba1c47fa
+size 1728

assets/demo-image-for-i23d/for-vae-reconstruction/Animals/0/10017/2/bbox.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d737f763e225f89a78191e3c629fda149b280aa4213432e06baa12433f647707
+size 192

assets/demo-image-for-i23d/for-vae-reconstruction/Animals/0/10017/2/c.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:958056548219cde6d6dfcf0307c23a2b00e9b6dde9d2bae29081962890c849e5
+size 1728

assets/demo-image-for-i23d/for-vae-reconstruction/Animals/0/10017/3/bbox.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:94e5b0e456eae3296b484125aa707c4c1e5d03d2e017a710f1cb7960c1e5f8ed
+size 192

assets/demo-image-for-i23d/for-vae-reconstruction/Animals/0/10017/3/c.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1b6285c92b810b127fa3780da3648788a9f58d2ddabd581847d802a06841c445
+size 1728

assets/demo-image-for-i23d/for-vae-reconstruction/Animals/0/10031/1/bbox.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4027472cdb5b3e2a4c668294960aa56861dbd886c3fe9b645e5cdf1fc3d60a89
+size 192

assets/demo-image-for-i23d/for-vae-reconstruction/Animals/0/10031/1/c.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ede83c74f20954e83d758e15d5917dba92525bc1f6b48162510061df48c1500f
+size 1728

assets/demo-image-for-i23d/for-vae-reconstruction/Animals/0/10031/2/bbox.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:90b64de4f85562412008570b6210422801b05aed07f4ef9e239e6b957407cc0f
+size 192

assets/demo-image-for-i23d/for-vae-reconstruction/Animals/0/10031/2/c.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2dbc68942442a97cf9a42723ee69c314e6970028a541b420e3a462fc0ba747f6
+size 1728

assets/demo-image-for-i23d/for-vae-reconstruction/Animals/0/10031/3/bbox.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:282630473f295afd5284813ee7b073d0ea3ce544445fb53d9007fbe3051689d2
+size 192

assets/demo-image-for-i23d/for-vae-reconstruction/Animals/0/10031/3/c.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d367864f87085597f0164ad56fd83447695b4ea4245aa0d98fc67a7845839fad
+size 1728

assets/demo-image-for-i23d/for-vae-reconstruction/Animals/0/10050/1/bbox.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7117c23b4cc5c07630a56f65414f8b623ed67886096d96f1533ab0055a4d07b7
+size 192

assets/demo-image-for-i23d/for-vae-reconstruction/Animals/0/10050/1/c.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a259ee690ac870ee8c9a5d528034a910d81c29bf592facde1c783fa62f5b5ac0
+size 1728

assets/demo-image-for-i23d/for-vae-reconstruction/Animals/0/10050/2/bbox.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9377f4782c5c8c54442ec752cb64e26b20efdc68dc532ae89fc464a98416a026
+size 192

assets/demo-image-for-i23d/for-vae-reconstruction/Animals/0/10050/2/c.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:99e603f1f429a309d5d88ff5b98996ac59904aad7a1b17b864baef188cb7b7cd
+size 1728

assets/demo-image-for-i23d/for-vae-reconstruction/Animals/0/10050/3/bbox.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cd9f62a5075c9acc4bcb9ee0b1983cfece27f9ce3dbffed241cf5cf02ac8f6df
+size 192

assets/demo-image-for-i23d/for-vae-reconstruction/Animals/0/10050/3/c.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ce7b47f5a7b3d88b1693d66af79b61aa17b424c816bb2ab1290a0ac98f59ea63
+size 1728

assets/demo-image-for-i23d/for-vae-reconstruction/Animals/0/10075/1/bbox.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b6bd3438f01f92caf7bb23014a5ec02b5304a3f3768ac97e408791415f222c7c
+size 192

assets/demo-image-for-i23d/for-vae-reconstruction/Animals/0/10075/1/c.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:33030a0a3a980c9b7831e924c25b573b5d8a25961c8392971c03b06a82613c34
+size 1728

assets/demo-image-for-i23d/for-vae-reconstruction/Animals/0/10075/2/bbox.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2f04ec25dba691e4a57fb81d4d7442b6bc729d722ee79ef51b0fe86c8566388a
+size 192

assets/demo-image-for-i23d/for-vae-reconstruction/Animals/0/10075/2/c.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ca81a811d44f2e735a34dadb75dd261832e23f6abd79636a1fe675c1803f3d65
+size 1728

assets/demo-image-for-i23d/for-vae-reconstruction/Animals/0/10075/3/bbox.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e38e13c22c0cd9ffac7dc334f72c9df3dbc15948a74e73e14d807fe771018d31
+size 192

assets/demo-image-for-i23d/for-vae-reconstruction/Animals/0/10075/3/c.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fcddb1c415599aab1e841bb4a460344427e5367490cb0d30966e96954aa0744f
+size 1728

assets/demo-image-for-i23d/for-vae-reconstruction/Animals/0/10118/1/bbox.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:19db7cc1d5064687a6a4144aa486a0658d94f29ed7dacae745b1d2f5e8fa47b6
+size 192

assets/demo-image-for-i23d/for-vae-reconstruction/Animals/0/10118/1/c.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cc7da40d8b8e29ed695f04a5bd86ddefe7b700cb9e647ba56b9a2e36f1f054b6
+size 1728

assets/demo-image-for-i23d/for-vae-reconstruction/Animals/0/10118/2/bbox.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:42a4e4eeb24d8a262ed13521ba6e91de49474e216d61fd079ae7148ba689d172
+size 192

assets/demo-image-for-i23d/for-vae-reconstruction/Animals/0/10118/2/c.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dc68dfa937cc6c012dfb3692f7502ba74156fa8e7a74738025e95c555726389e
+size 1728

assets/demo-image-for-i23d/for-vae-reconstruction/Animals/0/10118/3/bbox.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:abadb3c081d001607df41dd6572ccab30fec5a4d78e007cd27f88402cc6653d6
+size 192

assets/demo-image-for-i23d/for-vae-reconstruction/Animals/0/10118/3/c.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6d9c402a2a9a15c935a19ee400f56ff2bbee8fc56bc628f39b6a98a162baeca0
+size 1728

assets/demo-image-for-i23d/for-vae-reconstruction/Animals/0/10120/1/bbox.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:44213c91567fc2754967c0b4a941499165a4f48bf781bb0bca5ee042ce0276ef
+size 192

assets/demo-image-for-i23d/for-vae-reconstruction/Animals/0/10120/1/c.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bf128e21469f06af06597702f3816ba49caae49d60bd0cb283c9f140accc858d
+size 1728

assets/demo-image-for-i23d/for-vae-reconstruction/Animals/0/10120/2/bbox.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:19bdaa8de6263b6b28f45ea1078dc38684e5f299b72047af2f1253b2f89faac9
+size 192

assets/demo-image-for-i23d/for-vae-reconstruction/Animals/0/10120/2/c.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:91ebae020c0250173070293dc8a13e0aa43c15b08caf2b65b46a9f732f897829
+size 1728

assets/demo-image-for-i23d/for-vae-reconstruction/Animals/0/10120/3/bbox.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d5ce368b8c1cb9b653e383e0b6bb01420e104a746e9848a855aab7d1a266ea22
+size 192

assets/demo-image-for-i23d/for-vae-reconstruction/Animals/0/10120/3/c.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:83545b5bc970f348c3a6671f34164896a2ee1ac6b0141cb11c524ba3c98ad91f
+size 1728

assets/demo-image-for-i23d/for-vae-reconstruction/Animals/0/10955/1/bbox.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a3f4d41f65fbf790b9917522b4beef851a3fcd53bcc3e511c222eed13ecc2cb6
+size 192

assets/demo-image-for-i23d/for-vae-reconstruction/Animals/0/10955/1/c.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4e03a7e43edf3103b1752d43d959393bfd2cd6d3e2afc696772ffd5fa8cc68f5
+size 1728

assets/demo-image-for-i23d/for-vae-reconstruction/Animals/0/10955/2/bbox.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7065dd2259f73730dc729431af3cb198afd211fbf38ac48e27b7a57f2065d5cd
+size 192

assets/demo-image-for-i23d/for-vae-reconstruction/Animals/0/10955/2/c.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:993b9d849057f3a058449294db9cbda0f35a96e4f457d9f48b2de858e582e7db
+size 1728

assets/demo-image-for-i23d/for-vae-reconstruction/Animals/0/10955/3/bbox.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4605afccf0fd6fa49071150b7f9643ac79ce88c573df860b087cb8c1c1c2004f
+size 192

assets/demo-image-for-i23d/for-vae-reconstruction/Animals/0/10955/3/c.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:82a0e55ff4254fddcaa615b2dd12bc3d8942f76111313f588b8fc805d4698273
+size 1728

assets/demo-image-for-i23d/for-vae-reconstruction/Animals/0/12926/1/bbox.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:86d5b3723ac6398e61c71048af194f75406ad567aaa2a13e7e688247361a4198
+size 192

assets/demo-image-for-i23d/for-vae-reconstruction/Animals/0/12926/1/c.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d097407848b958efa6a87f564d83965b7d3a5cca506b0abc98285ca1ce2952c2
+size 1728

assets/demo-image-for-i23d/for-vae-reconstruction/Animals/0/12926/2/bbox.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fb8f52f9c3eea11ead7474e50aca3b55ca26a8332c83513735cf46992968e9c2
+size 192

assets/demo-image-for-i23d/for-vae-reconstruction/Animals/0/12926/2/c.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d89b2b9d82c1c5a06b25e4bde11aaa0e90d54a72a5e2b3e7077a0c2b8c89104d
+size 1728

assets/demo-image-for-i23d/for-vae-reconstruction/Animals/0/12926/3/bbox.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3c0e86d091d7dfb76e260040b51315277b61f9ab9f52b6a81154db0c2c9ca18c
+size 192

assets/demo-image-for-i23d/for-vae-reconstruction/Animals/0/12926/3/c.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:17430f723d12e4117b6af3e6b8bf1e68e5f0d4cfead403c1eb583f217909fc96
+size 1728