@LocalOptima it also adapts well when used together with other Loras

2.2 blows the legendary 2.1 away, and that's saying something. Gotta have a good workflow, gotta REALLY understand strengths and stuff when you're doing complex stuff, prompt game has to be on point, as with any model, and I don't think many people know this, but you can actually use 2.1 LoRas* with wan. It gives mixed results, some work better than others, but the output quality of 2.2 and its prompt adherance AND its general motion natural-ness smokes 2.1 so bad... It smokes all of them. Even the paid. Just gotta be patient with the setup, get the good workflow going, as in, get one that works, then tweak/change nodes as needed til you get a workflow you want. For example, my current workflow is scalable to generate literally endless video with no degradation. This can also be done on 2.1, but since the initial output quality of 2.2 is so much cleaner and consistent, a 2 step process can get me the material I need to generate a 2 hour video if I wanted to spend 5 dollars on renting a gpu overnight, lol. My current setup generates a 5 second video in 2-3 minutes on an a40, its not "faster" than 2.1, until you look at the deets. The deets show us that wan2.1 with lightning 2.2 generates higher resolution video with cleaner images AND cleaner motion in the same amount of time as 2.1, consistently, with better prompt adherence. If you tweak your settings/nodes/prompts, you can get 2.2 to REALLY crank out vids REALLY fast. And they look stupid good. The hybrid setup (one or two passes on high noise, 2-3 passes on high noise lightning, and 2-3 passes on low noise lightning) literally generate the best looking, smoothest, most realistic, consistent A.I. videos I've ever seen in under 4 minutes.

And for those of you who are wondering, creating endless video (reliably) is a 2 step storyboard process. You first generate a single video that has all of the elements you want within the 5 second time frame. When you generate it, save it to png files. So for example. You want to make a video of two meerkats fighting, that lasts 5 minutes. (who DOESN'T want to see this, all day) So you get your setup, throw any needed loras in, do a couple of 1 second test runs at 480 to make sure proof of concept works, then turn on your "save" function on your output node, bump up the strength and time and render out a few videos and pick the one you like the best. THAT IMAGE SET is what you're going to use as a "first frame" for each part of your generation. If you do it like this you don't need* colormatch (still recommend for good measure, but not needed). Then you take those images into your multiple step long video generation workflow, a good one, (mine is bullet proof at this point) and you set it up where you use the first image as the first frame on the first generator, then you pick an image from the set that you want the action to  flow to for your "last frame" on the first generator, and that same frame is ALSO plugged into your "first frame" of your second generator. pick your last frame for your second generator, that's your first frame for your third. Pick your last frame for your third, that's your fourth. IN THE EVENT that you have a LORA or a prompt that is going to DRASTICALLY change the image (the meerkats get muddy) INSTEAD of using a last frame on that generator, just use the first frame, and then use the "final frame selector" to pull ITS last frame to feed to the next video. It can get tricky keeping up with that, especially at the 2-3 minute mark, HOWEVER, if you plan your stuff accordingly, you can do a run of 5-6 generators, until the image is changed "too much" and then use that final frame to generate a new story board batch, or, use that last frame with qwen image edit or flux kontext to generate your storyboard and then continue as you did before with a new generation set. ONce all of the generators have generated all of their frames, and assuming its all in one setup, then all you gotta do is let the compiler compile all of the frames into one long vid. IF ITS MULTIPLE passes, as in, you gotta occasionally stop to build a new storyboard, then all of your images can be named accordingly, "gen1, gen2," etc, and then you can just load them into video combine and render your 5 minute meerkat fight candy. (or porn if you're gross)

@LocalOptima Ah thanks a lot!

both

you are missing the fact that the download is a zip file with both LoRAs in it.

Sounds like a shit workflow

@legolover nope, only change is that this lora is making asses way too big.

Look at the gallery, you're the only one having this problem. And if other loras are also changing the bodies then it's a you problem

I'd post an actual example of what you're getting so others can see.

It's the training data. This is most likely trained on just the insertion. You'll have to extend with a missionary/sex lora using the last frame.

@Slowmoe I'm sorry, a what loRA? All I see is "*******/*** lora". Not sure what happened there.

@stellarcosanostra0 Weird. Shouldn't be censored. Miss ionary / s e x lora.

@Slowmoe Gotcha. I'll try something like that out. Thanks.

@LocalOptima Looking forward to seeing whatever that is. You're hitting everything out of the park.

@LocalOptima Awesome :). Thank you! Can't wait to test it.

@LocalOptima Hell yeah, love your style, homie, don't take no shiz from the trolls! Thanks for what you do

@LocalOptima I cant prove this sadly(I say that but i think some of the images created with the models are still in my posts?) but I had a ton of models and workflows that I removed when Civitai went ott on censorship, I moved them to Tensor, then tensor made civitai's form of censorship look like nothing, I doubt any of my models are publicly accessible now but feel free to look my user name up on there (its the same) and you will see nothing is locked behind a pay wall and all is downloadable (as these things should be)

I have also provided tons of time coding for an open source game engine (legend of mir3 zircon files) along with massive tutorials for it. Until large language models made it irrelevent I had a free app to assist dungeon masters for D&D on the android, 2 years of coding, always free. I have helped on countless projects 90% for free, I have done paid work too, but only when the project is not based on open source ideals.....

just because you do not see what I share here does not mean I have not done so.

Patreon for donations is fair, not something I would do but yeah I get it. Locking workflows behind patreon is a kick in the teeth to the open source community "hey I took your work, made a cool thing and now charging money for it". You go on believing what you like of me, but dont paint me like I am freeloading. creations made with free open source projects should NEVER be behind a paywall, its a fundamental belief. Its a slippery slope.

there are 1 billion wan workflows online already lol

@LocalOptima ah yes because thats directly the same.

@NorfolkDave Just get the Kijai Workflows theres literally everything u need

Pick one of the many videos down below. Drag it into comfy. WAN is as easy as comfy gets, you don't even need the majority of extra crap in there. But if you really want to learn, follow along with a youtube tutorial and build a wf tailored to your setup. I stopped getting errors when I did this. If your system can run full models you can eliminate most of the optimizations and tradeoffs that cause issues, or conversely you can ditch them by building a wf with small quants, ggufs. Once you have the skeleton of WAN built, which is really only dozen or so nodes, nothing really, you can adding the extra stuff one by one. Any LLM can help you when you get stuck. That's my honest opinion, because that's how I learned it. Or you can try to debug/adapt a bloated broken workflow of dubious origin. That's more headache than learning, imo.

you can try add " the processed meat hot dog looks like a real penis"

Noticing some aiming problems in the examples too where the penis is put in too low or otherwise not into the vaginal canal. Anyone have a good negative prompt for that?

got a solution? same issue here, and its the case with all his great loras

best bet is gonna be to train a character lora.. im doing anime and with a character lora it works fine.. so i figure if it can do cartoons it can do a black man :D

Stuff like this can sometimes happen if you mention man twice.. like "Man appears and man inserts penis in her pussy", just as the simplest explanation. Try to see if your prompt has anything that might make "it" misunderstand that there are two men.

But yeah sometimes I feel like the generations just get .. "bugged" that once they do one thing, no matter how much you try to prompt it out, it just keeps happening. Had the same thing in one project with a nipple appearing through a shirt before the shirt opens and it just kept happening again and again and again, until finally it stopped and I don't even know what I did differently. (Had random seeds set, so it wasn't the seed)

@sarcasm83334 That fixed it, thanks! I tested a few prompts and used "man" once then "he/him/his" after that and now I don't have some random stranger popping into the frame. It's kinda funny that you can use "woman" 20 times and only get 1 but just use "man" twice and suddenly it's a threesome.

Latent homosexuality.

way I did it clip 1 - setup , clip 2 insertion at 1, clip 3+4 no insertion, dr34mjob instead, clip 5 insertion at 0.25 + cum loras, clip 6 only cum loras

just use wan2.2 experimental nsfw lora: about 1.4 strength high and 1.0 strength low.
link: https://civitai.com/models/1307155/wan-22-experimental-wan-general-nsfw-model
and sex goes further

Yeah, it's beyond obnoxious.

The problem is that if you upload them as individual files, then you have to make two entries and it's shit because you can only cite 8 resources on a post. Your complaint should be directed at how poorly this website is designed. Putting two large files in a compressed archive is a sane and logical decision.

Really a problem. Can someone help? How you can use Lora in this case?

@murkotinfaxelo780 are you trolling? You guys are aware that a zipped file...can be...unzipped?

@SaburoDio They're complaining that they can't brainlessly use whatever extension on ComfyUI they're using because it's shittily coded and can't handle archives.

I'm assuming you ment that the Metadata/Model-Info doesn't get downloaded, because ComfyUi lacks the info to get it from CivitAI. If you use the Lora-Manager you can right-click on the Lora-Card and Re-Link to CivitAi.

@SaburoDio like i can't understand why people complain about a zipped file? can't they just unzip it? Genuinely cannot understand this

@BekkaT8 Then you don't understand how things work automatically. Of course we can unzip it, that's easy. The issue is, civitai matches the data based on file hash. Once we unzip it, then there's no matching link on civitai's site. By providing a safetensors file, it lets automated tools use civitai's API to get all the file data from the file hash.

@phazei ah i see now, i have much more to learn, ty

@phazei I am seriously curious though... why not use comfy portable?

@SaburoDio what's that have to do with anything? I use Stability Matrix, which manages my models, ComfyUI is just installed in a python env, which I use for lots of other AI stuff.

@phazei ok but how many files are you using and do you realise YOU are having a more complicated approach than most normal users?

So you are complaining about driving a car on a bike path

@SaburoDio I mean, that's just the definition of me, I have a more complicated everything and it's a pita. But in general things should just work to standards, and then you don't need to think about it. People complained and now there are fewer people using zips. It might be more complicated, but it's for the purpose of more flexibility. And the hash lookup's are there to make everyone's life easier.

I get this problem too when the length goes too long but I find 81 frames is the sweet spot. 121 will always do this. I've experienced this issue with WAN 2.2--it tends to ping-pong with long batch sizes. It sucks but for now I just chain smaller batches together. Don't give up yet. If you keep having it, perhaps share your settings.