========================================================
Illustrious version
I had always wanted to train this lora for pony, but for some reason I just cant make it work; until IL was published. Anyway, the rate of getting a perfect image generation (a good image with right anatomy, poses, atmosphere, and outfit) is still quite low, around 20%; depending on the characters in prompt, some characters get higher rate, some gets lower.
IL ckpt can generate good fight scene too, this lora can stabilize results when multiple loras were used.
I suggest finding a result where postures of both characters are ok first, followed by inpainting each character out afterward; or whatever method works for you, such as regional prompt and other methods. Sometimes you don't need to Inpaint, the generation just came out right; but if inpaint is needed, I suggest reading my article on how to work with multiple characters/loras: My humble beginner guide to work with multiple Loras for complex composition. | Civitai It is heavily outdated, but the concept itself remains unchanged.
I might publish an update for this lora in future if i can find a way to improve the successful rate, as always have fun and go break some eggs!
========================================================
SD 1.5 version
V1.1 face punch
I tried multiple methods, seems like separating the attacks into different loras worked the best, so I would update them 1 by 1 here. this is a very difficult one, SD doesn't understand who's punching and who's getting punched.
Use regional prompt, latent couple, controlnet everything, you gonna need them. My advice will be dropping the facepunch lora to about 0.3 or turn it off when you are inpainting the character who's punching; increase the facepunch lora to about 0.8 for the character getting punched.
I had made 2 trigger words for this lora, 1st is yomama_face_focus, 2nd one is yomama_half_body, now with the half body one you should get the bigger frame done, i suggest using that in txt2img, use face focus on the character being punched in inapint.
It is a very difficult one, but with some patients and time, you will get the results. I will upload some openpose or depth for help soon.
As always, have fun and go break some eggs!
V1
A lora to generate base image of people fighting, in a more dynamic pose. The lora itself works well with 0.6-1.0 depends on how heavy you want the style to be.
Works well with "Punching" in prompt; "holding weapon" could work, but SD hates weapon as always, you can pair this lora with "Unlimited blade works" by @FallenIncursio .
You can use inpaint, regional prompt, latent couple to get what you want with different characters.
As always have fun and go break some eggs!
Description
FAQ
Comments (14)
nice
Gotta give this one a go! Looks cool, thanks for sharing <3
thanks! please let me know the results :D
Hi, what did you use for different characters? I'm confident to Latent Couple, so I think that's not what you used.
It's a bit outdated, but basically inpaint them individually. You can consider regional prompt too.
@yomama123556778 hardcore! lol, I didn't know that was an option, I'll check it out! Thanks for the response, when I have some spare time, I'll make a review!
@IndolentCat no problem, its not easy task tho, be prepared :D
It's crazy to me that your loras aren't more popular. Each one I've tried has been absolute fire.
Aww thanks for the support as always, love your work!
@yomama123556778 Do you think it's possible to train a lora with people actually making contact? Swords clashing, blocking, faces getting punched and so on?
@Flexability ah, i do plan on blocking and face punching, in fact I was collecting data for that already; unfortunately i got stuck re-training the kamehameha update, so you gonna have to wait a bit lol.
Sword clashing can be tricky, you know how SD goes with weapon, but i will see what i can do.
@Flexability oh btw, when you say, blocking, do you mean block with bare hands or with a shield?
@yomama123556778 I mean hand to hand, leg, etc
@Flexability aha, i was on the wrong direction, glad i asked lmao






