Flux:
I often use 3D character training and generation to evaluate the performance of models. Through this Flux model, I found that the performance of Flux.Dev in this evaluation still lags significantly behind Dalle-3 from two years ago. The biggest issue is the inability to reproduce complex body language.
---
Anna and elsa have the same training intensity, what can I say, anna is just shy, you just need to repeat the call in the prompt, something like "anna, anna,(anna:1.4)," also works for the other arousal words.
We have to admit that for SD1.5, the only way to get some kind of complex effect is to train or use a lot of time to tweak the prompt, and we're really looking forward to a version of stable diffusion with comprehension similar to Dalle3.
Description
FAQ
Comments (4)
why "Dall-E 3"?
What do you mean?
means the lora was trained off of Dall-E 3 generations
@cardpepe @ghostpaint Ah! That's pretty the same I'm trying to do, me too! : )
Have a look at my character https://civitai.com/models/289338/dalle-girl-scout-ambassador
I managed I trick to get a consistency, read here:
https://civitai.com/models/289338/dalle-girl-scout-ambassador?dialog=commentThread&commentId=346569
You can also have a look at the original dataset I got from DALLE.
https://mega.nz/file/Oo911CDQ#HXm5D_VJ5pR-mDU5BpzuXm9dNMWSHm1nowX4bryCed0
( Unfortunately the LORA, not made by me, was not so resemblant to original.... )
@settima_ai yeah, there's an SDXL Pepe lora that uses the same-ish technique as well. Using the generations from an AI model that has a way greater grasp on characters, prompt understanding, and general ability as a guider for a lesser-able model is probably the best strategy, unless there's already tons of pictures for a character in the style someone wants.