v1 is something I had to train, here it is, I have trained a lot, maybe it is overtrained because sometimes the “deep” actions needed to be searched (spin seeds).
Trained on 256 res dataset, captioned with joycaption sharpest frame.
The settings:
resolutions = [256] dataset.toml
x202 clips 3sec ( some are 2sec) 256x256 24fps,
learning rate 7e-5
num_repeat 2
v1.1
v1
v0.2
v0.1
A POV_DE3PTHR0AT of a latina woman wearing a {construction worker hat|police hat|firefighter hat|soldier hat|helicopter pilot hat} and very thick lips, wearing a light-colored top, she is looking up kneeling in front of a {white|black} man giving him a deepthroat blowjob. The woman has her mouth open. The angle is showing the woman from the top and the man is standing with his legs and feet visible. She moves her head up and down on the cock. The background is a black textured surface, possibly a bed or couch. The woman's neck is visible, and she appears to be enjoying the act. h3adm0ve, f4cesma5h, l1pssma5h, lip5onp3lvis, p3nissw4ll0w, whit3eye5, ell3r0ll.A POV_DE3PTHR0AT of an asian woman with long black hair, wearing a black top, giving a deepthroat blowjob to a man. The background is dark. The woman moves her head up and down, swallowing the penis. The lighting is dim, and the focus is on the woman's face and the man's penis.A POV_DE3PTHR0AT of a blonde woman with long hair styled in two braids, giving a deepthroat blowjob to a man. The background shows a room with wooden walls and green plants hanging on the wall. The woman is seen moving her head up and down while deepthroating the cock.***** This LoRA is shared without warranties and with the condition that it is used in a lawful and responsible way. I do not support or take responsibility for illegal, harmful, or harassing uses. By downloading or using it, you accept that you are solely responsible for how it is used. *****
Description
v1 is something I had to train, here it is, I have trained a lot, maybe it is overtrained because sometimes the “deep” actions needed to be searched (spin seeds).
Trained on 256 res dataset, captioned with joycaption sharpest frame.
FAQ
Comments (17)
Text to Video OR Images to video???
It's hunyuan, so both
How do you get more motion? Also how do you not lose motion when upscaling? Sorry I am quite new to this.
Upscaling works by upscaling each frame individually, motion should not be impacted at all.
amazing, thank you.
also, how do you guys create such long videos? is there any workflow that can solve that for smaller cards?
lower resolutions will allow longer videos, then you can upscale and interpolate the frames...
initial sample at 12fps, interpolate to 24fps. if you don't mind a little 'slo-mo' affect... i can get 8-10 sec videos at 480x720.
@az420 thanks!
@tedbiv good to know, thanks! is there a workflow for interpolation you recommend?
@olivereads38255 the workflow embedded in my HV LORA showcase videos is a pretty good one :) i.e. https://civitai.com/images/67179993
@az420 thanks, it's very different from what I'm used to, but I'll give it a shot!
Sadly😢, v1 is compared to v0.2, the flexibility is greatly reduced and the character LoRA barely works.
Yeah it might be overtrained, I'll test later on my end, ty for the feedback I used 5e-05 learning rate, I guess that might be the issue...
im testing now, i do see the character i loaded on the lora stack but i guess it loses movement and she doesnt want to goo as deep xD ,, if youre using ultra allinone workflow you can pass the character lora on the upscaler exclusive loras for getting the best results... also i dont know if having the fast lora active can mess with the other loras, since is on the "all" block
ok, Fast Lora wont affect the output, even sharing 1.00 strg with other lora in "all" blocks, im getting the character in the output.. what worflow you're using?
@K3NK First of all, thank you for your hard work.
I use allinone ultra 1.5v. LoRA stack is character 1.0(double block) + this LoRA 1.0(all blocks).
Based on my subjective impressions based on several personal LoRAs I use, character reproducibility is 80% (90%+ when upscaled) for v0.2. Movement is also good, but it depends on the seed.
But for v1 is about 10%. Almost similar female always appears and has a little bit of the characteristics of character LoRA. For example, if character LoRA has a dot on the corner of her mouth, the resulting female will not look like character LoRA at all, but will have a dot on her mouth 😂.
I tried adjusting the strength of the LoRA stack, but there was no significant change.
But it's okay, v1 is not useful to me, but v1 can be useful to someone. just like v0.2 is useful to me.
@dowefa5467894 im testing another training run i made, with 7e-05 LR, as i did with the previous ones, just this one has 256 resolution, i might upload the v1.1 today