Comparison to GLIDE-small (as shown in Figure 9 in the paper)

[benuri]

Figure 9 in the paper shows results for GLIDE-small which has been trained on the same data as GLIDE-full but has only 300M (like GLIDE-filtered).

When testing captions from Figure 9, like a hedgehog using a calculator, the results from GLIDE-small are clearly better than the results generated by https://replicate.com/afiaka87/laionide-v3.

What's the difference between GLIDE-small and Laionide-v3, besides the dataset?

• Do both models have exactly the same code?
• Do both models have trained with the same training code for the same number of iterations?