pov titfuck, a woman with enormous breasts with a penis in the cleavage, penis is center frame, and the breasts are moving up and down to rub on the penisDescription
FAQ
Comments (30)
Works great, thanks! 👍
looks great , i wish you will add some more samples, how long did you trained? im having difficulties with deepthroat t2v,, thats why i never published one t2v in wan.. wan2.2 is no better in this regard,, you could even say that is censored af.. i mean,, seems like hunyuan had penises on its main official model.. and wan does not..
The most important part was training at 720 resolution. 2e-5 through 5e-5 , 45 frames all work fine. I wasted a lot of time trying to train something at 320 and 512.
vfx_aido you have a 4090? which buckets are you using?
im really having a frustratin experience with wan and blowjobs T2V..
i guess youre notr using automagic?
K3NK it is possible. Try 720p and all the same aspect ratio
Kiefstorm i disabled enable_ar_bucket idk if that will allow it to start.. i cant even start 512 with 3 buckets.. so what are the best settings for not OOM nor using blockswap and being able to train at 720?.. does the number of videos matter? again i got oom , this sucks so bad..
https://pastebin.com/HpV9Gavp - training settings, I trained it on a 16GB vram
@vfx_ai lora rank 16?
K3NK 32rank
vfx_ai managed to start training at 720 rank32 with 16 frame bucket, multiple_overlapping, so thats 1sec vids.. just wat i wanted to try, training a high noise one with low res, and the low one at higher res with shorter vids.. i hope it works.. idk what to do now.. searching for more deepthroat vids is getting me tired..
K3NK this is all the 2.2 5B model, sorry
vfx_ai yeah i know, no worries 🤙
K3NK we believe in you K3NK, don't give up!
Is WAN 2.2 14b unusable?
no. i tried and the blocks dont load due too size mismatch. the generation runs, but the Lora has no effect.
I haven't used 5B, didn't realize it could create something like the sample. Does it gen faster or just less VRAM?
you gotta keep to the 720p resolution or else it goes haywire, but it is decently fast for the HD quality
It's much smaller and much faster than the 14B. It needs high resolution to make good things.
vfx_ai Do you have a workflow example? I'd like to try out the 5b
How does 5B 2.2 compare to 14B 2.1? This seems pretty good. What's everyone's experiences? And how about loras on 5B? Do they give as good results as loras on 2.1?
Haven't used it yet, but from what I've read, not enough info on good settings, and tries have led to bad results.
It has been used as an upscale model, so there's that, and it outputs at 24fps instead of the 16 fps of 14b. In my opinion (haven't used it yet), not as good as even old 2.1, but given it can run on lower hardware and Loras like this being better than 1.3b, it has potential to fill in the gaps a bit.
lets say that it allows me to do 120 frame generations on 960x1280 resolution with 12gb vram. Its not possible on 14b memory wise and not possible quality wise on 1.3b
Pretty annoying that Civit isn't adding a 5B model category.
Seems like they don't really careuch about keeping up with new models.
This Loras works well, but the resolution is limited. Other resolutions hide the woman's head.
prompt for it, and it will appear
bro can you share workflow pls/
kumarkishank959811 I use the default comfyui workflow from the comfyui examples page
maybe use image resize node along with pull camera back prompt/input spacious image
Actually amazing how smoothly and well this runs! probably the smoothest/best LoRA ive seen so far for this. Is this the true power of 5b instead of 14b? hm lol
awesome results with human models, but not good with anime xd