May be your lightning/lightx2v lora setup. They're still super finnicky. I'm just using the wan 2.1 lightx2v LORA at weight 1.5 and getting at least usable results. Have an article in my profile that explains my setup. It's not great, but it does work for now until we get usable 2.2 LORAs, etc.

with Wan2.2, lightx should be 2.5 high, 1.0 low (and, of course, I mean the working 2.1 version of lightX). 3+3 iterations is great. But you must be careful about how many steps the high thinks it is using, and also the noise setting of the second stage (low)- that top enable/disable noise setting).

Also the sampler/scheduler makes a difference. What does NOT make a difference is the render rez- if you are getting colour/washed out errors, try to find another workflow!

It is very easy for a workflow to malfunction down to all the things I mentioned above- which means you either need the confidence to experiment, or you need to try other workflows.

Wan2.2 has more 'black magic' settings' than ever did wan2.1

K3NK That's what works best- why? Who Knows. There's a lot of mistaken 'common knowledge' with Wan2.2- hence the complaint from psspss... A lot of bad workflows, repeated again and again here and elsewhere. If the lightX crew really knew what they are doing, their 2.2 version of LightX would work (better) than 2.1- but it does not.

My workflow has the high stage think it is doing only 3 iterations- in other words high 'completes a render'- then the low pass thinks it is completing the final 3 iterations of 6. The low pass enables 'add noise' (which doesn't happen on my i2v workflow). Why? Again I'm not certain, but it works.

There will be people incorrectly adjusting some of these settings, or maybe even trying to use res_2s and bong_tangent- notorious for bad output when incorrectly used (and amazing when correctly used- usually in t2i).

Oh, and my workflow was not my creation- it was simply the best I found using A-B testing. But it gave perfect results first time with your LoRA- and with other t2v wan2.2 LoRAs - what works works!

blobby99 BRILLIANT. Trying your 2.5:1 trick made me realize the low pass loras weren't even being applied in my SwarmUI attempts. Checked for updates and sure enough, Swarm had an update today to fix low pass lora not running.

Which also means every video I've tried (t2v OR i2v) so far has only had the high pass lora applied... no wonder I was getting wonky results like women trying to EAT the dicks.... 🤣🤣 Ah, frontier software....

whatsthisaithing How do you apply the Lora in SwamUI to the low model / refiner? Still via prompt?

marqs89 I've got a couple of articles in my profile that show I how setup wan 2.2 high and low models/loras in Swarm. For i2v you use <video> and <videoswap>, for t2v you use <base> and <refiner>. Pretty straightforward, but article shows all the settings. That said, I've switched to just working in Comfy directly at this point.

whatsthisaithing Thank you, i will have a look! But yeah, same, rn Comfy seems "easier" to work with (sadly) :D

K3NK hope to see this lora for i2v soon

K3NK Just realized in my SwarmUI workflow, the low pass LORAs (control AND yours) weren't being applied at all. Explains my weird results. Giving it another whirl this afternoon.

It works fine on I2V, use the dick lora with lower strength in addition and you're good

Any prompt por i2v?

@ThatHurts Use this instead, it's I2V: https://civitai.com/models/1874811/ultimate-deepthroat-i2v-wan22-video-lora-k3nk
Use the prompts in the examples, it creates the dick without being in the starting image.
You can also mix it with this Lora here of course, just reduce the strength a bit. Also works great in combination with the oral insertion Lora (ultimate + insertion)