Im back and going through backlog of loras to get some space back in my backup hardrive.
V3 - trained 32 frames 320x240 This res works best for base
Can do 32-128 frames tested follows the schedule fairly well 10fps-15fps
Settings - Lora 0.6 strength
Context length 16
Context stride 2
Context overlap 2
Positive schedule used
"0.0:0.4" :"smile",
"0.4" :"(cumshot, ejaculation:1.2)",
"0.4:1.0" :"(cum on face:1.1), wide-eyed, surprised"
Neg schedule is not needed but can help
"0.0:0.4" :"(cumshot, cum on face, cum:1.5)",
"0.4:1.0" :""
Ip adapter works quite well with photos cut at shoulders. You will need to play with the input image sometimes to get the right perspective
High ip adapter at 0.6 works well.
Upscale should be done with 2x upscale at steps 7-10 only or you will lose the animation
Models used in examples
Base
City edge toon mix
mysticmerth
Upscale - photon, hijklmix
Use of a cum lora is best 0.4 strength
https://civarchive.com/models/143625/cum-ii
Animate Diff Motion Lora - Cumshot
V2 - 350 steps trained - 32 frames 320x240 tall
This is the original V1 training data but this was trained on 32 frames instead of the original 16. This allows for 64 frame (more untested) videos at 15fps instead of the original 32 frames at 8-12 fps.
320x240 tall works, 512x256 works. I did not test much else. I want to try to add more frames to this to test.
Strength can be 0.8-1.1
Multivals both at 1.
Helper Lora available. Its not needed but may help allow stronger ip ref/lora use. I tested this very little but 0.3 strength helped hold the hand well. Your choice if you use it.
V1 alt - 16 frames 320x320 350 trained steps
Motion module influence 0.9-1 (effect multival)
This setting can fix the hand holding penis blur, 1.0 tends to distort, 0.8-9 works depending on your models
Strength of the lora 0.8-1.1
Can combine with other V1 lora at 0.4-0.8 strength each.
Not better than the other version, but works well enough and combines with the other.
Ip adapter is really hard to use with this most of the time.
Ref video has tongue out can prompt that away
32 frames works, i could not get more to not distort
stride of 2 works best but 1 can work
Dont loop, as ending is not the same as start.
You dont need a cum lora but it helps, the cum shows up either way but depending on your model is how it looks
Models used for testing
If using simple prompt travel
1girl, 1boy, 22 year old woman, realistic, (holding penis in front of face)
"0.0:0.4" :"",
"0.4:0.5" :"(cumshot, ejaculation:1.1)",
"0.5:1.0" :"(cum on face:1.1)"
Frames trained 16 320x320 200 steps recommended strength 1
If you can use 320x320 for base, and upscale. if not, try to keep it square.
The results come out at 12 fps IMO... can slow or speed that up, frame interpolate from there
Tried many models and sorry to say many did not work well. Cum is something some models really struggle with, test your model to see how it even does images like this first to see what you are working with.
Models i found worked really well for the base step, others worked ok for upscale
Mysticmerth V10
kawaiirealisticanime v4
cityedgetoonmix 3.12
This is the lora used in all examples set at 0.4 strength can go up to get more messy.
https://civarchive.com/models/143625/cum-ii
Unlike my other loras this requires a scheduled prompt for proper timing of things. Can work without, but it will start with cum on face
Because of this, i recommend using my prompt travel flow to get the best use of this.
Example prompt at bottom
I also recommend not using IP adaptor or loras till you get it running without. Then ease into them, this is very touchy. The animation works a lot better alone, but light influences are easier.
IP adaptor is hard to use with this without messing everything up. You can try, only luck I had was IP LIGHT at 0.4 strength linear
Upscaling can be tough, i use tile with soft scaled weights 0.875 to hold the cum shot animation, can try without it, it depends mostly on the model you are using to upscale
Steps 8-12
Tested 16,32 frames, I did not want to write the schedule for 64 so that's untested.
32 frames is its sweet spot, 16 barely gives time to do anything
Stride 1 for 16, 2 for 32, 4 for 64+ Overlap of your choice.
Example Prompt
Pre text
1girl, 1boy, (holding penis in hand:1.1)
Schedule
"0" :"clean face, (cum on face:-1)",
"5" :"clean face, (cum on face:-1)",
"6" :"(cumshot on face, ejaculation, squirt:1.1)",
"12" :"(cum shot on face, cum on face, cum in hair, open mouth:1.1)",
"16" :"(cum dripping, cum on face, open mouth, cum in hair:1.2)",
"26" :"(cum dripping, cum in hair:1.2)",
"32" :"(cum dripping, cum in hair, smile:1.2)"
App Text
, (long straight black Afro:1.2), dark skin, African American
Have Fun :)
Description
FAQ
Comments (2)
Received the error: "AnimateDiff - WARNING - No motion module detected, falling back to the original forward." when attempting to use. I was able to successfully use 5 other motion Loras, not sure why this one wasn't being identified. Tried re-downloading as well. Could just be me.
Amazing work. the thing is i have little knowledge about comfyui so can you add a simple workflow here to make animations without all the extra stuff like upscaler and controlnet ?