Tesla announces launch of universal AI fully autonomous driving solution
1051
English
English
For all friends who love creativity and need picture materials - whether you are a designer, a content creator, or just curious about AI painting - an exciting news is here! AI company HiDream.ai officially opens the source of their latest literary and artistic images (generate images based on text) big model: HiDream-I1.
This powerful model with 17 billion parameters (17B) not only generates amazing high-quality pictures for you in just a few seconds, but is also very friendly in use – completely free, open source, and even commercially available!
Simply put, HiDream-I1 is like a skilled and obedient AI painter who can help you turn the words and ideas in your mind into vivid images. It has several core advantages that make it easy to use:
Stunning picture quality and varied style: Whether you want realistic photos, cute cartoon images, or imaginative art paintings, HiDream-I1 can easily control it.
I am particularly "obedient" and understand what you think: Have you ever been worried about AI's "drawing the wrong questions"? HiDream-I1 is extremely good at understanding and following your text description (hint words), surpassing all other open source models in authoritative GenEval and DPG benchmarks.
Truly free, fully open: HiDream-I1 is open sourced with a very loose MIT license. It can be downloaded, used, modified or even studied freely without any hidden fees or restrictions.
Bold commercial use without worries: Using the images generated by HiDream-I1, you can freely use them in personal projects, scientific research, or even commercial products or marketing activities!
To meet the needs of different users, HiDream-I1 offers multiple versions, including the full version, the development version and the fast version. Each version is optimized based on the inference step and computing resources, and is suitable for different generation scenarios:
Full version : Provides the highest quality image generation with 50 inference steps. Suitable for scenes that require the finest image quality.
Download link
Development Edition : A good balance between efficiency and effect is achieved, with 28 inference steps.
Download link
Speed version : The number of inference steps is 16, suitable for scenarios that need to be generated in real time.
Download link
HiDream-I1 adopts a DiT model of Hybrid Expert Architecture (MoE), combining dual-stream MMDiT blocks and single-stream DiT blocks to efficiently allocate computing resources through dynamic routing mechanisms. In addition, the model integrates a variety of advanced text encoders, such as OpenCLIP ViT-bigG, OpenAI CLIP ViT-L, T5-XXL and Llama-3.1-8B-Instruct, which greatly improves semantic comprehension capabilities and generates images that are more in line with expectations.
If you want to experience the HiDream-I1 model, you can access its GitHub repository and follow the detailed installation and usage guide. You can also download the corresponding model version directly from Hugging Face and choose the version that best suits your needs.
Tip: Flash Attention is required to use this model, and it is recommended to use CUDA version 12.4. Make sure you have sufficient network connection and disk space, especially the Llama-3.1-8B-Instruct component that the model depends on.
Official product website: https://vivago.ai/
Resource website: https://www.tkj.ai/ai-models/hidream-i1-new
GitHub repository: https://github.com/HiDream-ai/HiDream-I1
Hugging Face Model Repository:
