CompVis Community

university

AI & ML interests

None defined yet.

Recent Activity

compvis-community's activity

morgan 
posted an update 6 months ago
view post
Post
1300
Llama 3.1 405B Instruct beats GPT-4o on MixEval-Hard

Just ran MixEval for 405B, Sonnet-3.5 and 4o, with 405B landing right between the other two at 66.19

The GPT-4o result of 64.7 replicated locally but Sonnet-3.5 actually scored 70.25/69.45 in my replications 🤔 Still well ahead of the other 2 though.

Sammple of 1 of the eval calls here: https://wandb.ai/morgan/MixEval/weave/calls/07b05ae2-2ef5-4525-98a6-c59963b76fe1

Quick auto-logging tracing for openai-compatible clients and many more here: https://wandb.github.io/weave/quickstart/

multimodalart 
posted an update 6 months ago
multimodalart 
posted an update 8 months ago
view post
Post
26503
The first open Stable Diffusion 3-like architecture model is JUST out 💣 - but it is not SD3! 🤔

It is Tencent-Hunyuan/HunyuanDiT by Tencent, a 1.5B parameter DiT (diffusion transformer) text-to-image model 🖼️✨, trained with multi-lingual CLIP + multi-lingual T5 text-encoders for english 🤝 chinese understanding

Try it out by yourself here ▶️ https://huggingface.co/spaces/multimodalart/HunyuanDiT
(a bit too slow as the model is chunky and the research code isn't super optimized for inference speed yet)

In the paper they claim to be SOTA open source based on human preference evaluation!