January 14th.Zhiputoday announced a jointHuaweiOpen SourceNew GenerationImage Generation Model GLM-ImageThe model is based on the Atlas 800T A2 device and the MindSpore AI framework to complete the entire process from data to trainingIT'S THE FIRST SOTA MULTI-MODEL TO COMPLETE FULL TRAINING ON A NATIONAL CHIP.
GLM-Image uses the self-innovated "Self-Return + Proliferation Decoding" hybrid structure, which is an autonomous, innovative, self-repeated, self-demolition-demolator complexJoint image generation with language models achieved.
1AI with GLM-Image core highlights as follows:
- Structural innovation, oriented towards "cognitive generation" technology exploration: using a "self-regression + proliferation encoder" hybrid structure, taking into account global command understanding and local detailOVERCOMING THE CHALLENGES OF CREATING KNOWLEDGE-INTENSIVE SCENES SUCH AS POSTERS, PPT AND COPUTU, a step towards exploring a new generation of “knowledge + reasoning” cognitive generation models, represented by Nano Banana Pro。
- The first SOTA model to complete full-scale training in a nationally produced chip: the model ' s self-return structure base is based on the Rotation Atlas 800T A2 device and the MindSpore AI framework, which completes the full process construction from pre-processing of data to large-scale training, and validates the feasibility of training forward models on the National Product Total Calculator Base。
- Text Rendering Open Source SOTA: First on the CVTTG-2K (complex visual text generation) and LongText-Bench (long text rendering) listsHe's very good at the Chinese word generation.
- PRICE-FOR-MONEY VERSUS SPEED OPTIMIZATION: API CALL MODEGenerating a picture cost $ 0.1, the speed optimization version is about to be updated。
According to the official spectrograph, GLM-Image was able to adapt itself to multiple resolution, by improving the Tokenizer strategy, and originals supported the task of generating an arbitrary proportion of images from 1024 x 1024 to 2048 x 2048, without retraining。
GLM-Image reached in an authoritative list of wordsOPEN SOURCE SOTA HORIZONTAL.
GLM-Image tested the following in the actual complex graphic tasks:
Scene I: Cope Art
The GLM-Image is better at drawing schematics and schematics of science with complex logical processes and narratives。
Scene II: Dog Pictures
GLM-Image is able to keep styles and subjects consistent and to guarantee the accuracy of multiple text creations when generating multi-grand drawings such as electric graphs and comics。
Scene three: social media graphic cover
GLM-Image is used to create complex images, such as social media cover and content, to make your creation more free。
Site IV: Business poster
GLM-Image is able to generate design-sensitive, text-embedded holiday posters and business outreach maps。
Scene Five: Live Photography
In addition to text replicating, GLM-Image is also very good at generating images, pets, landscapes, still objects of all kinds and sizes。
1AI with GLM-Image experience and open source addresses as follows:
- online experience: https://bigmodel.cn/trialcenter/modeltrial/image
- API access: https://docs.bigmodel.cn/cn/guide/models/image-gender/glm-image
- GitHub: https://github.com/zai-org/GLM-Image
- Hugging Face: https://huggingface.co/zai-org/GLM-Image
- ZhipuAI/GLM-Image