At the recently held World Artificial Intelligence Conference, Xu Li, chairman and CEO of SenseTime, brought some recent system capability upgrades of the new SenseTime model. Let us take a look at the changes in this upgrade.
SenseChat 2.0 fully upgrades the large language model capabilities
Model basic ability improvement
The basic ability of the model has been greatly improved
New regional languages (support Arabic, Cantonese, etc.)
Break through the input length limit of large language models
new service features
Newly added knowledge fusion interface,
Optimization based on knowledge base capabilities
model hallucination
LLM model system
lS small model version 1.0
lXL large model version 2.0
SenseMirage 3.0 fully upgrades the Vincent diagram large model
Lightweight Painting Customization
Everyone can easily (just drag and drop) and quickly (within 10 minutes) fine-tune the model, and customize their own personal generated AI.
intelligent description mode
Using the large language model as a bridge, the "complex description" is simplified into "a simple prompt word", and a picture with rich content can be generated.
Photographic Image Reproduction
Comparable to a professional photographer's level of filming, the details and elements of the image are truly realized
With the self-developed large-scale model (7 billion parameters as the base), the imagination of AI painting is endless.
AI code assistant: performance is improved in an all-round way, and the accuracy exceeds GPT3.5
AI FOR AI New 28th Law: Code = 80% AI Generation + 20% Artificial
HumanEval test set pass rate 48.7% (GPT 3.5 48.1%)
Code generation speed per token: 25ms
Reasoning throughput: 1.9kTokens per second
1.0
Code completion, code extension, code translation, code refactoring, code correction, comment generation code, complexity analysis, test case generation
2.0
Multiple rounds of dialogue, adding comments to the code, explaining the code, and filling in the middle of the text...
Code writing efficiency increased by 78%
Multiple rounds of dialogue, better interactive experience
Performance improvement, higher coding efficiency
Chinese enhancement, more in line with the habits of Chinese developers
Ronin: comprehensive improvement of digital human performance
Flow generation: simultaneous input and output, double the efficiency
Multilingual accuracy improvement: English, Korean, Japanese, Arabic, etc. increase the accuracy of more than 30% languages
Cinematic Clarity: 4K HD Digital Human Video
Complex environment support: distance range, face angle greatly improved
Qiongyu 2.0 digital twin application platform algorithm upgrade
Centimeter level reconstruction accuracy:
Outdoor: Accuracy of 5 cm (1%) per 10,000 square meters
Indoor: Accuracy of 1 cm (1.5%) per 1000 square meters
Efficient real-time rendering
The reconstruction efficiency increased by 20%, and the construction time of 100 square kilometers was 38 hours
Rendering performance is improved by 50%, and the 1080p resolution rendering of a single card can reach real-time
Rich Asset Features
Support output model surface depth
Support synchronous generation of 3D Mesh model
Support weather environment simulation
Support NeRF semantic analysis
Support CGCS2000 national geodetic coordinate information alignment
Support ground-air data fusion
Gewu 2.0
Algorithm upgrade
More detailed, millimeter level, more material support - high reflection, more light and shadow control
Commodity 3D digitization
Automobiles, jewelry, cosmetics, home decoration, home appliances, clothing and other types can be displayed in 3D digital form.