Total runs:
5.7M
Run Growth:
1.5M
Growth Rate:
25.55%
BAAI / AltDiffusion-m18
Model's Last Updated: October 02 2023
Introduction of AltDiffusion-m18
Model Details of AltDiffusion-m18
AltDiffusion
| 名称 Name | 任务 Task | 语言 Language(s) | 模型 Model | Github |
|---|---|---|---|---|
| AltDiffusion-m18 | 多模态 Multimodal | Multilingual | Stable Diffusion | FlagAI |
模型信息
AltDiffusion-m18 是一种基于@StableDiffusion 的多语言文本图像生成模型。该模型由 Stability AI 和@BAAI FlagAI 团队合作完成（FlagAI 是 LF AI & Data Foundation 的沙盒阶段项目）。AltDiffusion-m18目前支持 18 种语言，包含：英语、中文、日语、泰语、韩语、印地语、乌克兰语、阿拉伯语、土耳其语、越南语、波兰语、荷兰语、葡萄牙语、意大利语、西班牙语、德语、法语和俄语。
AltDiffusion-m18 is a multilingual text-image generation model built on @StableDiffusion. This model is a collaboration between Stability AI & @BAAI FlagAI team (FlagAI is a sandbox-stage project of LF AI & Data Foundation). AltDiffusion-m18 currently supports 18 languages, including English, Chinese, Japanese, Thai, Korean, Hindi, Ukrainian, Arabic, Turkish, Vietnamese, Polish, Dutch, Portuguese, Italian, Spanish, German, French, and Russian.
训练方法
如图1，所示训练分为两个阶段：概念对齐阶段和效果提升阶段。我们首先替换使用多语言CLIP AltCLIP-m18替换掉原始SD的OpenCLIP, 之后冻住AltCLIP的参数。在第一阶段中，使用256*256的图片分辨率，训练Unet中CrossAttention层的k,v矩阵进行文图的概念对齐。在第二阶段中，使用512*512的图片分辨率，训练Unet的所有参数进行生成效果的提升。
As shown in Figure 1, the training process consists of two stages: concept alignment and quality improvement. We first replaced the original OpenCLIP in SD with the multilingual CLIP AltCLIP-m18 and froze its parameters. In the first stage, we trained the k,v matrices in the CrossAttention layer of the Unet model to align the concepts between text and image using 256*256 image resolution. In the second stage, we trained all the parameters in the Unet model to improve the generation performance using 512*512 image resolution.
数据使用
在第一阶段中，我们使用 LAION 5B 中的LAION 5B-en(2.32B) 和 过滤的18语言 LAION 5B-multi（1.8B）数据进行训练。在第二阶段中，我们使用 LAION Aesthetics V1 中的LAION Aesthetics V1-en(52M) 和 过滤的18语言 LAION Aesthetics V1-multi(46M)数据进行训练。
In the first stage, we trained the model using LAION 5B-en(2.32B) from LAION 5B and filtered LAION 5B-multi(1.8B) data for the 18 languages. In the second stage, we trained the model using LAION Aesthetics V1-en(52M) from LAION Aesthetics V1 and filtered LAION Aesthetics V1-multi(46M) data for the 18 languages.
训练细节
优化器：AdamW
学习率：1e-4 并带有10k步的warmup
显卡：64 张 NVIDIA A100-SXM4-40GB 第一阶段，从SD v2.1 512-base-ema开始，以batch size 3072在256*256的分辨率上使用64张A100训练330k步，耗时8天；第二阶段，从第一阶段330k的checkpoint开始，以batch size 3840在512*512的分辨率上使用64张A100训练270k步，耗时7天。然后，基于270k的checkpoint随机丢掉10%的文本进行150k步的classifier-free guidance训练，耗时4天。
The first stage involved using the SD v2.1 512-base-ema checkpoint to initialize all parameters except for the language model, with a batch size of 3072 and a resolution of 256x256 for training on LAION2B en and LAION2Bmulti for 330k steps over approximately 8 days. In the second stage, training began from the 330k step checkpoint, with a batch size of 3840 on LAION Aesthetics V1-en and V1-multi, and training for 270k steps with a resolution of 512x512, taking around 7 days. Training then continued from the 270k step checkpoint for another 150k steps, with 10% of the text randomly discarded for classifierfree guidance learning, taking approximately 4 days. The teacher model of AltCLIP is OpenCLIP ViT-H-14(version is ”laion2b s32b b79k”). The pretrained Stable Diffusion checkpoint we used is SD v2.1 512-base-ema. We also use Xformer and Efficient Attention to save memory use and speed up training. The decay of EMA is 0.9999.
效果展示
18语言效果
中文效果
长图效果
参考
--Stability AI: https://stability.ai/
--FlagAI: https://github.com/FlagAI-Open/FlagAI
--Stable Diffusion: https://huggingface.co/spaces/stabilityai/stable-diffusion
模型参数量/Number of Model Parameters
| 模块名称 Module Name | 参数量 Number of Parameters |
|---|---|
| AutoEncoder | 83.7M |
| Unet | 866M |
| AltCLIP-m18 TextEncoder | 1.19B |
引用/Citation
Please cite our paper if you find it helpful :)
@misc{ye2023altdiffusion,
title={AltDiffusion: A Multilingual Text-to-Image Diffusion Model},
author={Fulong Ye and Guang Liu and Xinya Wu and Ledell Wu},
year={2023},
eprint={2308.09991},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
许可/License
该模型通过 CreativeML Open RAIL-M license 获得许可。作者对您生成的输出不主张任何权利，您可以自由使用它们并对它们的使用负责，不得违反本许可中的规定。该许可证禁止您分享任何违反任何法律、对他人造成伤害、传播任何可能造成伤害的个人信息、传播错误信息和针对弱势群体的任何内容。您可以出于商业目的修改和使用模型，但必须包含相同使用限制的副本。有关限制的完整列表，请 阅读许可证 。
The model is licensed with a CreativeML Open RAIL-M license . The authors claim no rights on the outputs you generate, you are free to use them and are accountable for their use which must not go against the provisions set in this license. The license forbids you from sharing any content that violates any laws, produce any harm to a person, disseminate any personal information that would be meant for harm, spread misinformation and target vulnerable groups. You can modify and use the model for commercial purposes, but a copy of the same use restrictions must be included. For the full list of restrictions please read the license .
Runs of BAAI AltDiffusion-m18 on huggingface.co
54
Total runs
-1
24-hour runs
19
3-day runs
20
7-day runs
35
30-day runs
More Information About AltDiffusion-m18 huggingface.co Model
Url of AltDiffusion-m18
AltDiffusion-m18 huggingface.co Url
Provider of AltDiffusion-m18 huggingface.co
Other API from BAAI
Updated: February 22 2024
Total runs:
2.2M
Run Growth:
-2.1M
Growth Rate:
-95.77%
Total runs:
1.6M
Run Growth:
-57.4K
Growth Rate:
-3.60%
Total runs:
1.2M
Run Growth:
643.7K
Growth Rate:
54.16%
Total runs:
966.7K
Run Growth:
342.4K
Growth Rate:
35.39%
Total runs:
961.0K
Run Growth:
-2.7K
Growth Rate:
-0.28%
Total runs:
954.2K
Run Growth:
637.5K
Growth Rate:
66.81%
Total runs:
759.4K
Run Growth:
141.1K
Growth Rate:
18.58%
Total runs:
152.3K
Run Growth:
-20.2K
Growth Rate:
-13.28%
Total runs:
139.5K
Run Growth:
44.8K
Growth Rate:
32.14%
Total runs:
131.4K
Run Growth:
16.9K
Growth Rate:
12.88%
Total runs:
25.7K
Run Growth:
1.4K
Growth Rate:
5.51%
Total runs:
12.4K
Run Growth:
4.6K
Growth Rate:
34.71%
Total runs:
7.5K
Run Growth:
-146
Growth Rate:
-1.95%
Total runs:
6.9K
Run Growth:
-11.1K
Growth Rate:
-84.72%
Total runs:
6.7K
Run Growth:
-2.6K
Growth Rate:
-37.67%
Total runs:
6.5K
Run Growth:
319
Growth Rate:
4.92%
Total runs:
5.5K
Run Growth:
-421
Growth Rate:
-7.64%
Total runs:
5.4K
Run Growth:
-699
Growth Rate:
-13.06%
Total runs:
5.3K
Run Growth:
-572
Growth Rate:
-10.83%
Total runs:
5.3K
Run Growth:
-599
Growth Rate:
-11.35%
Total runs:
5.0K
Run Growth:
2.6K
Growth Rate:
52.13%
Total runs:
4.3K
Run Growth:
1.5K
Growth Rate:
33.97%
Total runs:
3.8K
Run Growth:
-300
Growth Rate:
-7.98%
Total runs:
3.7K
Run Growth:
-400
Growth Rate:
-10.78%
Total runs:
2.8K
Run Growth:
196
Growth Rate:
7.14%
Total runs:
2.0K
Run Growth:
1.1K
Growth Rate:
54.31%
Total runs:
1.3K
Run Growth:
1.1K
Growth Rate:
88.53%
Total runs:
1.2K
Run Growth:
25
Growth Rate:
86.21%
Total runs:
664
Run Growth:
0
Growth Rate:
0.00%
Total runs:
644
Run Growth:
0
Growth Rate:
0.00%
Total runs:
476
Run Growth:
-633
Growth Rate:
-29.46%
Total runs:
355
Run Growth:
16
Growth Rate:
4.51%
Total runs:
310
Run Growth:
-148
Growth Rate:
-47.74%
Total runs:
199
Run Growth:
62
Growth Rate:
30.85%
Total runs:
133
Run Growth:
-161
Growth Rate:
-121.05%
Total runs:
97
Run Growth:
69
Growth Rate:
71.13%
Total runs:
64
Run Growth:
-95
Growth Rate:
-150.79%
Total runs:
53
Run Growth:
11
Growth Rate:
17.74%
Total runs:
41
Run Growth:
3
Growth Rate:
100.00%
Total runs:
28
Run Growth:
-440
Growth Rate:
-427.18%
Total runs:
27
Run Growth:
-955
Growth Rate:
-3537.04%
Total runs:
24
Run Growth:
1
Growth Rate:
100.00%
Total runs:
24
Run Growth:
3
Growth Rate:
100.00%
Total runs:
19
Run Growth:
-12
Growth Rate:
-63.16%