@c0ur4ge THanks, I still have to test this lora, I was curious because the examples are in slow motion and this is very common with speed lora.

@c0ur4ge what strenght do you recommand for lightning for H and L ? thanks

@stablediffusion2660 Start with 1/1 and adjust from there.

@ERA5ER @c0ur4ge You guys should check out my comment over here for another new idea for WAN 2.2 gens, it's the best thing I've tried so far, I posted a comment on this workflow: Wan 2.2 14B I2V Tests of Sampler Setting with and without speed LoRAs - v1.0 | Wan Video Workflows | Civitai

Do you have a good workflow that you slotted this into? Having trouble getting it working well with the default Wan2.2 I2V template.

@c0ur4ge I figured it out. I was an idiot and had both the high and low noise models using the Lora instead of only the high.

@sleeptightchamp196 can you share your workflow plese.

@c0ur4ge I understand. Will this help generations with prompts without the trigger words for motion/NSFWfying?

@c0ur4ge what other word could missionary, anal, blowjob, etc be mistaken for? just use those words. hint: theyre unique

Try it without the lightning loras on the high pass, those can definitely mess with motion.

@dixienormus935 I just tried. It's hard to say conclusively if its better or not, but in general i feel they are about the same using 4 steps on the high pass without lightning lora. Appreciate the suggestion though. I'll keep tinkering

EDIT: I tried both configurations with the same seed, and the results are almost identical, just tiny nuances that are different, but both have the same "stifled" motion, so i dont think its the high pass lora thats getting me. Perhaps this is as good as it gets and i'm just being picky 😂

@Foxbite you need to increase the number of steps for the high pass if you're not using the lightning LoRAs, try 10 steps instead. Otherwise, the issue is likely with your prompting; make sure you're highly descriptive and give a reason for your character to be acting quickly. That being said, even with ideal prompting I sometimes get output with the slowmo effect, but it's usually only about 10% of the time, then another 10% they're moving way too quickly. But 80% good results is good enough for me

Oh one other thing, how many tokens are in your prompt? You want to try and hit a sweet spot of around 256 tokens, beyond that I've had issues with attention which can definitely cause the slow motion effect. So you want to be descriptive, but not overly so.

@c0ur4ge cool, interested to see what you come up with, thanks for taking a look!

looking forward to it :)

@c0ur4ge Thanks for the insight! I'll give shorter clips a try. Yes I have messed up the timestep differences before and "mega cursed" sounds about right, or maybe "cronenberg horror".

It's a shame these models take so long to train on my basic bitch consumer GPU. With SDXL it's pretty easy to experiment with learning rates and datasets and such and bang out a few iterations in a day, but with Wan I can spend a week on a single model and not have anything to show for it.

Anyway incredible job on the Lora, really great stuff you've been putting out!!

@c0ur4ge lightspeed lora CFG 1.0 8 steps max (4 high/4 low). 4 steps (2 high/2 low)

@c0ur4ge I haven't tried Kijai's workflow, but grey screen is only when using the light lora and your lora together. With just the light lora I don't have the issue, and with just your lora and no light lora also not having it. Your lora should work with the default workflows so it doesn't make sense that Kijai's would work. That is why I offered the extra details... and I see other people reporting similar issues. Washoping for some other suggestions like Limono mentioned. I'll try what Limono said for the lightspeed though it wont help that the results even without lightspeed aren't as clear as other people's with light artifacting and stuff regardless of even doing 1024x1024 with 40 steps. It would be ideal if you had a workflow for this, too. Surprised you didn't link one in your post. Would be a good idea to consider with optimized sampler, CFG, etc. and notes.

yes trust the other guys, my wan gens were also bad at first , every little setting matters!

@c0ur4ge yet ot would be great if u could share a workflow for an example it has been generated with so people like me (noobies) could get into it to figure out more

@eugskb532 im looking for that as well. im using the stock comfyui wan 2.2 models and workflow, tried just change the loras but my images are distorted and grainy, cant see anything. must be the workflow isnt correct to work with this lora

@c0ur4ge I literally told you what my parameters, including sampler, were in my original post. I've tried the comfy default workflow, Kijai's, and as mentioned some of the ones below that other users used for their results. None of it worked. I've even tried it at various strengths and CFG mixes just in case that was related since there wasn't proper info for that provided. The reason we ask for a workflow is because none of these workflows work with it at the default settings meaning there is some setting in the workflow we're not using that it specifically needs. Ex. sampler you asked about, but I already told you what one I'm using... Euler. Does it only work with specific samplers?

@immauser blurry and distorted results are not about wrong loras. It's about wrong sampler/model settings. All of the loras which are confirmed compatible with WAN 2.2 works fine with the default comfyui template. Just play around with CFG or increase steps

@c0ur4ge Thank you for your reply. I tried adjusting the CFG, and now the control seems to have improved. In fact, I had never considered that the issue might be related to CFG before.

@c0ur4ge What is an average CFG you would recommend, I usually glide around 3.5, and if I want "more" prompt adhesion, I should raise it right? By how much?

@Simon_Belmont  i use cfg3.5 will have something wrong,i just use 2

High dictates motion. Low is detail. This is why you can use Wan 2.1 character loras in the low ksampler where they make the most Impact since they directly affect detail

More practically, the HIGH model is meant for the early generation steps, the LOW is meant for the later generation steps. It's a two stage workflow just like a two stage workflow used for image generation models

I'm gonna give you the answer you are looking for: in Wan 2.2 there's two "routes" a hign and low noise. Both are used, everytime, so you need both. And you place the low noise on the low route and the high noise lora on the high noise route, as this is how wan 2.2 (NOT 2.1) works.

@c0ur4ge I found the Kijai model on huggingface, but dont know how to download it. With wan2.1 i was able to download as a safetensor file. Do you know how to download it from huggingface?

Also any idea where to find a workflow?

@javbukkake wow youre an actual clown

.

The same thing is happening to mine.

@c0ur4ge that's what I done and IT WORKS! Beautiful!!!!!! Thanks for your help!

@IAmNotABot4Real Hey, would you be up for posting that working workflow json somewhere so we can piggyback off of your success? <3

@RickDripps sorry for the late response. Where can I upload the JSON file?

@IAmNotABot4Real  you can Create an Article and upload your json file. PLS

@550023981196 Okay, I will do that.

@550023981196 here is the article: https://civitai.com/articles/20342/nsfw-wan-comfyui-i2v-workflow

I made this one > https://github.com/brandschatzen1945/wan22_i2v_DR34ML4Y/blob/main/WAN_2LORAs_32FPS.json ; contains 2 wan loras (disable the one you dont want) & is quantstack & 32fps smoother video >> low ram > max effort ; origin : https://civitai.com/articles/17964/new-version-v20-of-torstens-low-vram-wan22-14b-i2v-simplified-workflow (from torsten) and adapted it.

can you share workflow please for i2v

@illumizoldyck808524 sorry, I use Wan2GP not Comfy https://github.com/deepbeepmeep/Wan2GP

@kunde2 no problem. thanks for trying to help out :)

You must be doing something right, because this lora is much easier to use than the other general NSFW loras on this site ;)

@c0ur4ge Thanks for the info!

@c0ur4ge Hype

you don't need special option, just add the LoRA with a normal LoRA loader after the model, and before the ModelSamplingSD3...

Or, add a power lora loader and replace the existing lora loader because the power lora loader lets you load more than one.

Or use Wan2GP

hi, u can have a good one looking into a channel on youtuber called "jockerai" and look for a video name "WAN 2.2 10x". It is a good one workflow and work good on the rtx 3060 12vram gpu

Yeah there is and I have one

@contatocutscuts733 what am i doing wrong? im very new and added the models getting very blurry results. i followed his steps and no luck