SupraLabs_It has been a while since our last text-to-image model experiments... the resultsimages were always blurry, damaged and not good.
We spent days in researching better options and tricks for a better model now. Let's get into it.
The old model
Here are some real samples - generated from our model. At guidance_scale=3.5 and steps=150. Everytime the same 12345 seed.
Prompt: A clock tower against a blue sky.
Prompt: a red car
Prompt: a mountain landscape with a lake
As you can clearly see, these results are REALLY bad. That has to be improved.
The new setup
Since we are an open-source research organization we will share some details about the new setup yet - even before the release.
| Part | OLD | NEW | Why? |
|---|---|---|---|
| Encoder | CLIP | Flan-T5-Base | Better image understanding and prompt following improvements. |
| Dataset | COCO | FLUX-Reason-6M | 1. Synthetic images are better for small t2i models; 2. COCO is not good at all for t2i; 3. All images in the FLUX-Reason-6M dataset already have a squared aspect ratio. |
| # of images | 83k | 900k | More data is better to not let the model overfit to certain pictures |
| Epochs | 200 | 2 | With more data, you should use much less epochs. |
| Total parameters | 40M | 100M | Text-To-Image is complicated and very hard. Give the model more room to think. |
Let's see how it will perform!
Total time and current progress
We did start the training of the model. At 256² image size. With 900k images. For two epochs. One single RTX 5060 Ti 16GB.
Stay tuned
This was it for today - stay tuned for the release maybe in some days!