apolinario (multimodal.art)@multimodalart
May 14, 2024
5 tweets
The first open Stable Diffusion 3-like architecture model is JUST out 
- but it is not SD3! 
It is HunyuanDiT by Tencent, a 1.5B parameter DiT (diffusion transformer) text-to-image model 
In the paper they claim to be SOTA open source! I'm working on a @Hugging Face demo as you read this so we can all vibe check
Model: huggingface.co/Tencent-Hunyua
GitHub: github.com/Tencent/Hunyua
Paper: tencent.github.io/HunyuanDiT/ass
- but it is not SD3!
It is HunyuanDiT by Tencent, a 1.5B parameter DiT (diffusion transformer) text-to-image model
In the paper they claim to be SOTA open source! I'm working on a @Hugging Face demo as you read this so we can all vibe check
Model: huggingface.co/Tencent-Hunyua
GitHub: github.com/Tencent/Hunyua
Paper: tencent.github.io/HunyuanDiT/ass
The model is focused on working very well in chinese, so they use as text encoders: a multi-lingual CLIP + multilingual T5
The model is also capable of multi-turn conversation and providing edits
And here are the evals reported on the paper - based on human evaluation, where they claim top of the open source pack!
Demo is out! x.com/multimodalart/
apolinario (multimodal.art)@multimodalart
May 14 24
View on Twitter
Demo for the first open SD3-like architecture model, HunyuanDiT @Hugging Face Spaces demo is out! 
First impressions:
- Image quality seems very good!
- Chunky and the research code isn't super optimized for inference speed (
@diffusers 
)
huggingface.co/spaces/multimo
First impressions:
- Image quality seems very good!
- Chunky and the research code isn't super optimized for inference speed ( @diffusers )
huggingface.co/spaces/multimo Important to compare as well: x.com/multimodalart/
apolinario (multimodal.art)@multimodalart
May 14 24
View on Twitter
Btw, while the HunyuanDiT model shares the DiT architecture with SD3, it is a smaller model 
Which, based on results I've gotten/seen so far, makes the results even more impressive!
Nonetheless, it defo doesn't beat SD3 across the board - hope we see the largest full SD3 release soon 
Which, based on results I've gotten/seen so far, makes the results even more impressive!
Nonetheless, it defo doesn't beat SD3 across the board - hope we see the largest full SD3 release soon 
apolinario (multimodal.art)
@multimodalart
ML for Art and Creativity, working @HuggingFace (apolinario@multimodal.art)
Missing some tweets in this thread? Or failed to load images or videos? You can try to .