Had trouble creating POV shots, so made this helper lora trained on 28 images.
It's not perfect, play around with the strengths and seeds, but better than native.
use one of these:
- high angle
- eye level
- low angle
at the end of the prompt you can mention what parts of your body are visible
example prompt:
mpov, first-person perspective pov video, i am sitting on the edge of the cliff, my legs are bent at the knees and my hands are resting on them, i am wearing a green sweater and jeans, i lower my hands to my sides and stand up. high angle, only my hands and legs are visible.
P.S: it can do NSFW
I have also attached a zip file with other epochs from training, if the effect is not strong enough/too strong - try a different epoch. If you find a better checkpoint, please share results here.
Check this post for a special ref2va use-case: https://civarchive.com/posts/31276080
For anyone who wants to donate, please direct it to the creator of Fizgig: https://buymeacoffee.com/lorasandlenses
Note for model creators:
This was created on Fizgig with 32 images over 100 epochs (6-49, 16 rank, other settings default, ref2va), it didn't get there completely, if anyone has a better config or training preset, please comment it down here.
Description
FAQ
Comments (20)
This is one of those loras you don't know you need until you see that it exist.
I had a lot of problems with male POV, the models always make it so that he is seated or lying on his back, why is the male POV always seated or lying on his back?
Your lora is just what we needed for these cases
And also when you wanted to be seated he is lying. Well that means even seated is not consistent. :-D But overall what helps is to point that again in summary and also was using (camera = POV) in summary. But maybe with this lora I would not need it so much.
This looks great, I love POV videos. Especially when there's a girl there or something...
Thanks. Am I assume that this can also be used as female POV by simply promoting it? If not, any plans on female POV? Thanks for your work.
i didn't want to post anything too risque but this lora is trained specifically on male pov scenarios, a different lora with female pov will be released soon
Nice! I have had this trouble too.
Appreciated. Just one important prompting question:
Are we referring to the camera as a subject? Or how do you tell it to move around, just like we are referring to the guy?
Please provide example prompts.
I forgot to add a sample prompt, here is an example
mpov, first-person perspective pov video, i am sitting on the edge of the cliff, my legs are bent at the knees and my hands are resting on them, i am wearing a green sweater and jeans, i lower my hands to my sides and stand up. high angle, only my hands and legs are visible.
one of the reasons i wanted to make this lora - was to make it easier to prompt. prompting for camera/viewer felt weird and wasn't simple.
@rmanb2 ok so it is truly first person from "my" pov? I will try. thanks.
@rmanb2 my first try failed. i ended up with the person supposed to be 'me' sitting in front of the camera and there was no POV to be found. I will keep trying.
@kronos1959777 try changing the strengths. ALSO - some loras like realism people, plaguekind realism and better motion seem to always nullify the pov effect for some reason.
So, if you are using any loras along with this, try disabling them.
and make sure you add the trigger in the beginning AND the camera angle and visible body parts at the end.
@rmanb2 oh k i didn't do the ending part.
@rmanb2 yea lower strength helped. i posted my example
Handsome guy, when will female POV Lora be released~！
coming soon....
Problems with POV?
[Shot 1] The camera is the POV of (S1), he walks towards ...
summary: The target video is the POV of <Subject 1> as they walk through the forest at night.
detailed_description:
[Shot 1] The camera is the POV of <Subject 1> (S1), he reaches into the bag and removes ...
Never mention 'man' or anything that couldn't be seen in a POV shot, just describe the actions in first person, no face etc. Its a pain in the ass to prompt a POV shot because any little hint will break out of it. You must use caps also just like in the official guides its 'POV' and not pov, yes its so fucking anal like that...
it is possible in native, yes - but as you said - not easy to prompt.
and i prefer to write my prompts by hand, not use an llm, so this should make it easier.
"male viewer" is the correct wording for man from pov perspective
@rmanb2 Yeah I don't use llm either and using an llm is lazy mode and often puts its own cringe into the script which i hate. Anyway the other poster is probably correct "male viewer" would probably work great since its telling the model explicitly. But if using reference model with reference you do need all the sections subject_definitions, summary, and etc. Other wise it will not only get confused the quality will suffer a lot I find. I've always just said its the 'POV of the <Subject N>, as he ...' and it mostly works, it only ever breaks out when you describe things that would require a different camera position to render it, for example 'The man walks towards ... ' it will usually generate the man's full body shot instead of moving the camera in pov.
Thanks for making this, its not neccessary but it helps with the prompt and stability.
Thanks for making this. Native H3 POV is basically broken, it either goes third-person or spawns a duplicate. This LoRA actually works.Tested it in ComfyUI with ref2va on a male POV scene. mpov + trigger words work fine, angle system makes sense.Please don't abandon it. 28 images is small, but the demand is real,so many people here are asking for exactly this. Female POV would be the logical next step. Not asking now, just saying there's an audience.
Thanks for filling a gap native H3 can't handle