ArthurZ HF staff commited on
Commit
1585eee
1 Parent(s): b54479d

Upload README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +53 -1
README.md CHANGED
@@ -4,4 +4,56 @@ tags:
4
  - pytorch
5
  - diffusers
6
  - unconditional-image-generation
7
- ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4
  - pytorch
5
  - diffusers
6
  - unconditional-image-generation
7
+ ---
8
+
9
+ # Denoising Diffusion Probabilistic Models (DDPM)
10
+
11
+ **Paper**: [Denoising Diffusion Probabilistic Models](https://arxiv.org/abs/2006.11239)
12
+
13
+ **Authors**: Jonathan Ho, Ajay Jain, Pieter Abbeel
14
+
15
+ **Abstract**:
16
+
17
+ *We present high quality image synthesis results using diffusion probabilistic models, a class of latent variable models inspired by considerations from nonequilibrium thermodynamics. Our best results are obtained by training on a weighted variational bound designed according to a novel connection between diffusion probabilistic models and denoising score matching with Langevin dynamics, and our models naturally admit a progressive lossy decompression scheme that can be interpreted as a generalization of autoregressive decoding. On the unconditional CIFAR10 dataset, we obtain an Inception score of 9.46 and a state-of-the-art FID score of 3.17. On 256x256 LSUN, we obtain sample quality similar to ProgressiveGAN.*
18
+
19
+ ## Inference
20
+
21
+ **DDPM** models can use *discrete noise schedulers* such as:
22
+
23
+ - [scheduling_ddpm](https://github.com/huggingface/diffusers/blob/main/src/diffusers/schedulers/scheduling_ddpm.py)
24
+ - [scheduling_ddim](https://github.com/huggingface/diffusers/blob/main/src/diffusers/schedulers/scheduling_ddim.py)
25
+ - [scheduling_pndm](https://github.com/huggingface/diffusers/blob/main/src/diffusers/schedulers/scheduling_pndm.py)
26
+
27
+ for inference. Note that while the *ddpm* scheduler yields the highest quality, it also takes the longest.
28
+ For a good trade-off between quality and inference speed you might want to consider the *ddim* or *pndm* schedulers instead.
29
+
30
+ See the following code:
31
+
32
+ ```python
33
+ # !pip install diffusers
34
+ from diffusers import DDPMPipeline, DDIMPipeline, PNDMPipeline
35
+
36
+ model_id = "google/ddpm-cifar10-32"
37
+
38
+ # load model and scheduler
39
+ ddpm = DDPMPipeline.from_pretrained(model_id) # you can replace DDPMPipeline with DDIMPipeline or PNDMPipeline for faster inference
40
+
41
+ # run pipeline in inference (sample random noise and denoise)
42
+ image = ddpm()["sample"]
43
+
44
+
45
+ # save image
46
+ image[0].save("ddpm_generated_image.png")
47
+ ```
48
+
49
+ For more in-detail information, please have a look at the [official inference example](_) # <- TODO(PVP) add link
50
+
51
+ ## Training
52
+
53
+ If you want to train your own model, please have a look at the [official training example]( ) # <- TODO(PVP) add link
54
+
55
+ ## Samples
56
+ 1. ![sample_1](https://huggingface.co/google/ddpm-church-256/resolve/main/generated_image_0.png)
57
+ 2. ![sample_2](https://huggingface.co/google/ddpm-church-256/resolve/main/generated_image_1.png)
58
+ 3. ![sample_3](https://huggingface.co/google/ddpm-church-256/resolve/main/generated_image_2.png)
59
+ 4. ![sample_4](https://huggingface.co/google/ddpm-church-256/resolve/main/generated_image_3.png)