Wan2.2 Another workflow that do NSFW. But now - with custom LORA for each step.
Details
Download Files
Model description
RED COLOR LETTERS: Unbypass all the bypassed positives to bypass an error.
Here, fixed. Must do that long ago.
Specs: 4070TiS 16gb+64gbDDR5+i7-14700. So, you may ask why I'm posting it here? Because, it uses gguf (you can chose your own light models) and it can do long videos AND it can apply differen LORA for each step. That's it. I searched something like that by myself, but...
ATTENTION! Saampler is custom and putted in ComfyUI-WanMoeKSampler which must be putted into custom_nodes (damn... do I really must do 5.1 version with standart node?.............................................................................................................................................................................................................................. I really thought you can manage that).
Inspiration:
/model/1829052?modelVersionId=2070152 -I stole from this almost everything.
/model/1847579/wan-22-i2v-helper-loras-without-killing-motion-or-lora-action-lightning-22-i2v-pusa-fusionx Kiefstorm: thanks for proper cumshots, I finaly believe it is possible.
Sampler:
and
I corrected it for my pleasure, so I put the corrected one in the archive, but you can forget about it and use default two. Also, in theory, different ModelSampling for High and Low can play different roles. But supposedly Boundary plays an even bigger role. I don't understand all this, so I corrected the original ComfyUI-WanMoeKSampler for myself so I could sleep peacefully (you can download original, or just two samlers like others put, or I put mine from the screenshot in the archive so you don't have to worry, just copy the folder "ComfyUI-WanMoeKSampler" to custom_nodes, the original nodes.py is there under the name nodes1.py). Well, and the remaining billion nodes will have to be installed somehow by yourself.
Oh, and they undress much good if you give them space and time (I mean - change my testing 49 seconds to 97 atleast).
Be aware to chose resolution in WanImageToVideo node, not in load&resize - i saw profit of resizing only in WanVideoWrapper workflows (WanVideo ImageToVideo Encode node throw my system in space), and I DON'T KNOW WHY it do that if same WanImageToVideo node do same thing and не выёбывается.
And thanks to this forking miracle: "wan2.2 GGUF Kijai longvideo.json" (also put it in archive, for education). I don't really remember where I got this darkflow from, but it worked purely on WanVideo nodes (which not compatible with normal JUST BECAUSE) and produced such disgusting quality with even BSOD that I bravely fought with it to understand how to rip out the LORA for each step prompt. And it was impossible, since the developers tried to do WanVideo nodes not like others somewhy. Guy, we use same models, why your nodes works different? Why we can't even chose sheduler (oh yes, you named sampler liker scheduler, and your "scheduler" do a crappy choise in terms of wan2.2).
I did it my own way. Maybe it won't be clear at first because of the connections, but the main nodes you will use should be visible - same as others.
Guys, I myself burst into Wan literally a month ago, you'll figure it out too.
Choose quantized models you have or want, experimenting with clip and clip_vision (they all the same). I wasn't intendet to do fastest wf. I want do more controllable.
Now you your own director.
P.S. Before that I was fapping to pictures from forge and fooooocus about a half a year, two years ago I tried your Comfy, but it seemed complicated and took a long time for gens on my hardware (imagine, on rtx2070 sd1 or sdxl was as fast as flux). And then suddenly, after years, after forge, a bit of invoke and foooocus for inpainting, I came across SwarmUI (thanks to StabilityMatrix - it shows options two point click install), and when I installed Swarm - it's basically the same comfy, and I was like - FUCK. Well, there was workflows right from the box. I tried. Also, Wan2.2 came. If you want to make videos, you'll have to work hard, I thought (or CogVideo from SMatrix - but thats crappiest shit I ever saw). So, everyone can handle everything, except for the eternal problem of finding the necessary nodes (in fact, you can also replace them with others you have).
Chose Loras you want, try mine or yours. Atleast it works. Atleast it can be a good start point, if you realise how it works. Also, be free to use dpmppppmdpdpdpfpdgppgpgpg_2m_sde and sgm_uniform. Or lcm and ddim_uniform (one guy on githum posted that that is the best thing he ever saw, but for me and my resolution - eyes always glitchy). Also no face consistency when face goes under the shirt or off the screen, but we all know that problem (just pompt "The WOMAN keep facing and looking at the camera." and hope it will). Best results - when the picture staring at you since the beginning. I'd was kinda shocked if I can be shocked when saw that.
Oh yeah, for the witnesses of the sacred boundary. 0.9 for 720p, 0.89 for 600p and 0.88 for 480p. Don't ask why. Also don't ask why HIGH SD3Sampler is set to 8.5 (it's not much different from the usual one), and LOV Noise is set to 4.5. I have no explanation for this. It didn't get worse - and that's okay. But I'm open to discussions on this matter.
I don't like upscaling with built in ways, so there is no that crappy plastic thing. But you may add, ofcourse. Also, you can lower resolution to 480p aaaaaaaand it wouldn't be worser. But I saw what 720p doing, and now I can't unsaw it.
upd 04:09: I have a lack of motivation to continue what I intendet to do, which is - atleast body consistency (if someone undresses something - it must be same proportions after), so carefully using loras for every different thing is a way, and I got them here, but to catch really working universal loras - that's hard (or they ruine body consistency, not telling about face, or every seed looks same (like oral insertion lora)).
I stopped my way on first-last frame, cause it is more predictable. Waiting for proper wan2.2 vace (yeah, tried WanVideo wf with fakewan or test_wan for 2.2 - works awefully somewhy for me). Checked really big wf's like a rocketship, complexed (whole venv must be redone for it, bunch of new nodes used for one time), for h100, but they all uses WanVideo nodes (that's why h100 needs), and result is like not worth it, that's why I don't like much WanVideo nodes (yeah, also no scheduler option).
I fixed some issues of this version of wf, added some better loras, even write nodes. But never published it. Cause no interesting example. I don't really wan't to post nsfw stuff, but all testings I did - is nsfw stuff. Lack of imagionation and motivation.