Comparing V3.0 and V2.5

Disclaimer: I don't have an artistic background nor a pro AI user. I'm just a normal guy who loves CF so much that I have to do a test with these 2 version for comparing. I don't speak english that much either so yeah, bear with me.

The test is carried out using the same seed, the same settings, tiled diffusion and tiled VAE for upscaling. The test concluded:

1/ Vertical (2 images)

2/ Horizontal (2 images)

3/ Lora (1 images) (Note: V3.0 has very wild way with lora, it won't listen to your lora at all)

4/ NSFW (1 images)

The first picture will be V3.0 and the second one will be V2.5.

My opinion:

1/ As someone in the discussion have already said: V3.0 use NAI clip so the geibooru tag and danbooru tag may not be effective. For V3.0, you might need to use a more "normalized" prompt. (Counterfeit's author also said this in his hugging face).

2/ V3.0 is very difficult to use with lora. Like. VERY DIFFICULT. Or it's just me, idk.

3/ Personally, I think V3.0 is more focused on people/character, while, V2.5 is likely proned to detailed background and person at the same time.

4/ I like both of 2 versions. V3.0 for easier prompt controlling and person-focused. V2.5 for more detailed background and lora. The moral of this test is just choose which version suits you the best.