The algorithms previously implemented for T2I-Zero were incorrect. They should be working much more stably now. See the previous result in the 'images' folder for an informal comparison between old and new. Some settings like the sampler are different (Euler a vs DPM++ SDE), so it isn't a real A vs B.
The algorithms previously implemented for T2I-Zero were incorrect. They should be working much more stably now. See the previous result in the 'images' folder for an informal comparison between old and new. Some settings like the sampler are different (Euler a vs DPM++ SDE), so it isn't a real A vs B.
Before:
After: