Files changed (1) hide show
  1. README.md +91 -36
README.md CHANGED
@@ -1,81 +1,136 @@
1
  ---
2
- license: creativeml-openrail-m
3
- datasets:
4
- - laion/laion400m
5
- tags:
6
- - stable-diffusion
7
- - stable-diffusion-diffusers
8
- - text-to-image
9
  language:
10
- - en
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
11
  ---
12
 
13
- # LDM3D model
14
 
15
- The LDM3D model was proposed in ["LDM3D: Latent Diffusion Model for 3D"](https://arxiv.org/abs/2305.10853) by Gabriela Ben Melech Stan, Diana Wofk, Scottie Fox, Alex Redden, Will Saxton, Jean Yu, Estelle Aflalo, Shao-Yen Tseng, Fabio Nonato, Matthias Muller, Vasudev Lal.
16
 
17
- LDM3D got accepted to [CVPRW'23]([https://aaai.org/Conferences/AAAI-23/](https://cvpr2023.thecvf.com/)).
18
 
19
- This new checkpoint use the depth as one channel compared to the previous version.
20
 
21
- ## Model description
22
 
23
  The abstract from the paper is the following:
24
  This research paper proposes a Latent Diffusion Model for 3D (LDM3D) that generates both image and depth map data from a given text prompt, allowing users to generate RGBD images from text prompts. The LDM3D model is fine-tuned on a dataset of tuples containing an RGB image, depth map and caption, and validated through extensive experiments. We also develop an application called DepthFusion, which uses the img2img pipeline to create immersive and interactive 360-degree-view experiences using TouchDesigner. This technology has the potential to transform a wide range of industries, from entertainment and gaming to architecture and design. Overall, this paper presents a significant contribution to the field of generative AI and computer vision, and showcases the potential of LDM3D and DepthFusion to revolutionize content creation and digital experiences.
25
 
26
  ![LDM3D overview](model_overview.png)
27
- <font size="2">LDM3D overview taken from [the original paper](https://arxiv.org/abs/2305.10853)</font>
28
 
29
- ## Intended uses
30
 
31
- You can use this model to generate RGB and depth map given a text prompt.
32
  A short video summarizing the approach can be found at [this url](https://t.ly/tdi2) and a VR demo can be found [here](https://www.youtube.com/watch?v=3hbUo-hwAs0).
33
- A demo is also accessible on [Spaces](https://huggingface.co/spaces/Intel/ldm3d)
34
 
35
- ### How to use
36
-
37
- Here is how to use this model to get the features of a given text in PyTorch:
38
  ```python
39
-
40
  from diffusers import StableDiffusionLDM3DPipeline
41
 
42
  pipe = StableDiffusionLDM3DPipeline.from_pretrained("Intel/ldm3d-4c")
43
- pipe.to("cuda")
44
 
 
 
45
 
46
- prompt ="A picture of some lemons on a table"
 
 
 
47
  name = "lemons"
48
 
49
  output = pipe(prompt)
50
  rgb_image, depth_image = output.rgb, output.depth
51
- rgb_image[0].save(name+"_ldm3d_rgb.jpg")
52
- depth_image[0].save(name+"_ldm3d_depth.png")
53
  ```
54
 
55
  This is the result:
56
 
57
  ![ldm3d_results](ldm3d_4c_results.png)
58
 
59
-
60
- ### Limitations and bias
61
-
62
- For the image generation, limitations and bias are the same as the ones from [Stable diffusion](https://huggingface.co/CompVis/stable-diffusion-v1-4#limitations)
63
- For the depth map generation, a first limitiation is that we are using DPT-large to produce the ground truth, hence, other limitations and bias are the same as the ones from [DPT](https://huggingface.co/Intel/dpt-large).
64
-
65
-
66
  ## Training data
67
 
68
  The LDM3D model was finetuned on a dataset constructed from a subset of the LAION-400M dataset, a large-scale image-caption dataset that contains over 400 million image-caption pairs.
69
 
70
  ### Finetuning
71
 
72
- The fine-tuning process comprises two stages. In the first stage, we train an autoencoder to generate a lower-dimensional, perceptually equivalent data representation. Subsequently, we fine-tune the diffusion model using the frozen autoencoder
73
 
74
  ## Evaluation results
75
 
76
- Please refer to Table 1 and Table2 from the [paper](https://arxiv.org/abs/2305.10853) for quantitative results.
77
- The figure below shows some qualitative results comparing our method with (Stable diffusion v1.4)[https://arxiv.org/pdf/2112.10752.pdf] and with (DPT-Large)[https://arxiv.org/pdf/2103.13413.pdf] for the depth maps
78
- ![qualitative results](qualitative_results.png)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
79
 
80
  ### BibTeX entry and citation info
81
  ```bibtex
 
1
  ---
 
 
 
 
 
 
 
2
  language:
3
+ - en
4
+ tags:
5
+ - stable-diffusion
6
+ - stable-diffusion-diffusers
7
+ - text-to-image
8
+ model-index:
9
+ - name: ldm3d
10
+ results:
11
+ - task:
12
+ name: Latent Diffusion Model for 3D-4C
13
+ type: latent-diffusion-model-for-3D-4C
14
+ dataset:
15
+ name: LAION-400M
16
+ type: laion/laion400m
17
+ metrics:
18
+ - name: FID
19
+ type: FID
20
+ value: 27.82
21
+ - name: IS
22
+ type: IS
23
+ value: 28.79
24
+ - name: CLIP
25
+ type: CLIP
26
+ value: 26.61
27
+ - name: AbsRel
28
+ type: AbsRel
29
+ value: 0.0911
30
+ - name: RMSE [m]
31
+ type: RMSE-m
32
+ value: 0.334
33
+ pipeline_tag: text-to-3d
34
+ license: creativeml-openrail-m
35
  ---
36
 
37
+ # LDM3D-4C model
38
 
39
+ The LDM3D model was proposed in the paper [LDM3D: Latent Diffusion Model for 3D](https://arxiv.org/abs/2305.10853.pdf), authored by Gabriela Ben Melech Stan, Diana Wofk, Scottie Fox, Alex Redden, Will Saxton, Jean Yu, Estelle Aflalo, Shao-Yen Tseng, Fabio Nonato, Matthias Muller, and Vasudev Lal.
40
 
41
+ LDM3D was accepted to the [IEEE / CVF Computer Vision and Pattern Recognition Conference (CVPR)](https://cvpr2023.thecvf.com/Conferences/2023) in 2023.
42
 
43
+ This new checkpoint uses the depth as one channel compared to the previous version.
44
 
45
+ ## Model details
46
 
47
  The abstract from the paper is the following:
48
  This research paper proposes a Latent Diffusion Model for 3D (LDM3D) that generates both image and depth map data from a given text prompt, allowing users to generate RGBD images from text prompts. The LDM3D model is fine-tuned on a dataset of tuples containing an RGB image, depth map and caption, and validated through extensive experiments. We also develop an application called DepthFusion, which uses the img2img pipeline to create immersive and interactive 360-degree-view experiences using TouchDesigner. This technology has the potential to transform a wide range of industries, from entertainment and gaming to architecture and design. Overall, this paper presents a significant contribution to the field of generative AI and computer vision, and showcases the potential of LDM3D and DepthFusion to revolutionize content creation and digital experiences.
49
 
50
  ![LDM3D overview](model_overview.png)
51
+ <font size="2">LDM3D overview taken from the [LDM3D paper](https://arxiv.org/abs/2305.10853).</font>
52
 
53
+ ## Usage
54
 
55
+ You can use this model to generate an RGB image and depth map given a text prompt.
56
  A short video summarizing the approach can be found at [this url](https://t.ly/tdi2) and a VR demo can be found [here](https://www.youtube.com/watch?v=3hbUo-hwAs0).
57
+ A demo is also accessible on [Spaces](https://huggingface.co/spaces/Intel/ldm3d).
58
 
59
+ Here is how to use this model to get the features of a given text in PyTorch on both a CPU and GPU architecture:
 
 
60
  ```python
 
61
  from diffusers import StableDiffusionLDM3DPipeline
62
 
63
  pipe = StableDiffusionLDM3DPipeline.from_pretrained("Intel/ldm3d-4c")
 
64
 
65
+ # On CPU
66
+ pipe.to("cpu")
67
 
68
+ # On GPU
69
+ pipe.to("cuda")
70
+
71
+ prompt = "A picture of some lemons on a table"
72
  name = "lemons"
73
 
74
  output = pipe(prompt)
75
  rgb_image, depth_image = output.rgb, output.depth
76
+ rgb_image[0].save(name+"_ldm3d_4c_rgb.jpg")
77
+ depth_image[0].save(name+"_ldm3d_4c_depth.png")
78
  ```
79
 
80
  This is the result:
81
 
82
  ![ldm3d_results](ldm3d_4c_results.png)
83
 
 
 
 
 
 
 
 
84
  ## Training data
85
 
86
  The LDM3D model was finetuned on a dataset constructed from a subset of the LAION-400M dataset, a large-scale image-caption dataset that contains over 400 million image-caption pairs.
87
 
88
  ### Finetuning
89
 
90
+ The fine-tuning process comprises two stages. In the first stage, we train an autoencoder to generate a lower-dimensional, perceptually equivalent data representation. Subsequently, we fine-tune the diffusion model using the frozen autoencoder.
91
 
92
  ## Evaluation results
93
 
94
+ ### Quantitative results
95
+
96
+ The table below shows the quantitative results of text-conditional image synthesis on the 512 x 512-sized MS-COCO dataset with 50 DDIM steps.
97
+
98
+ |Method |FID ↓|IS ↑ |CLIP ↑ |
99
+ |------------|-----|------------|------------|
100
+ |SD v1.4 |28.08|34.17 ± 0.76|26.13 ± 2.81|
101
+ |SD v1.5 |27.39|34.02 ± 0.79|26.13 ± 2.79|
102
+ |LDM3D (ours)|27.82|28.79 ± 0.49|26.61 ± 2.92|
103
+
104
+ Our model is on par with the Stable Diffusion models with the same number of parameters (1.06B). IS and CLIP similarity scores are averaged over 30k captions from the MS-COCO dataset.
105
+
106
+ The following table shows the evaluation results of depth evaluation comparing LDM3D and DPT-Large with respect to ZoeDepth-N that serves as a reference model.
107
+
108
+ |Method |AbsRel|RMSE [m]|
109
+ |---------|------|--------|
110
+ |LDM3D |0.0911|0.334 |
111
+ |DPT-Large|0.0779|0.297 |
112
+
113
+ The results shown above can be referenced in Table 1 and Table 2 of the [LDM3D paper](https://arxiv.org/abs/2305.10853.pdf).
114
+
115
+ ### Qualitative results
116
+
117
+ The figure below shows some qualitative results comparing our method with [Stable Diffusion v1.4](https://arxiv.org/pdf/2112.10752.pdf) and with [DPT-Large](https://arxiv.org/pdf/2103.13413.pdf) for the depth maps ![qualitative results](qualitative_results.png).
118
+
119
+ ## Ethical Considerations and Limitations
120
+
121
+ For image generation, the [Stable Diffusion](https://huggingface.co/CompVis/stable-diffusion-v1-4#limitations) limitations and biases apply. For depth map generation, a first limitiation is that we are using DPT-large to produce the ground truth, hence, other limitations and biases from [DPT](https://huggingface.co/Intel/dpt-large) are applicable.
122
+
123
+ ## Caveats and Recommendations
124
+
125
+ Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model.
126
+
127
+ Here are a couple of useful links to learn more about Intel's AI software:
128
+ * [Intel Extension for PyTorch](https://github.com/intel/intel-extension-for-pytorch)
129
+ * [Intel Neural Compressor](https://github.com/intel/neural-compressor)
130
+
131
+ ## Disclaimer
132
+
133
+ The license on this model does not constitute legal advice. We are not responsible for the actions of third parties who use this model. Please cosult an attorney before using this model for commercial purposes.
134
 
135
  ### BibTeX entry and citation info
136
  ```bibtex