They've already listed 5B on their todo list in the repo. But because its already pretty fast they're probably releasing it last.

Ada321 So in general, you can bet, right? I'm still working on the standard GUI and I'm new to this whole thing. I'm still practicing i2v generation. Is it true that i2v works better on a 2.2 WAN? Because on 2.1, very often the original image was broken or smeared.

Reading the official documentation is always a good start ;)
https://github.com/mcmonkeyprojects/SwarmUI/blob/master/docs/Video%20Model%20Support.md#wan-22

you have a Workflow handy? I seem to be struggling with this guy

xuadamux373 I use Academia's workflow. Check his channel on YT.

@Ada321 Exctrated to improve quality, speed or vram usage?

The self forcing lora just does that in the console. Has no effect on the video as far as I can tell, I mean beside it generally slowing down animation to that slow-mo effect vs just native WAN.

From my experience, the new Lightning Wan2.2 self-force lora doesn't do that, but the old one does that as @CAIdonni mentioned. Based on my limited understanding, it is due to the fact that these accelerator LoRAs are not trained with image keys. So when the console looks for an image key, it just isn't there. Luckily, it doesn't affect the process and it will still be used to generate the video.

TorstenTheNord i see, thx

Can you tell me more?

TeosKuzen yes like for example when I try to do ultimatedeepthroat lora for example it will play out like a boomorang video sometimes and the character just sucks the tip instead of going all the way and. I wondering if you are able to get loras to fully function the same as they would without this speed up lora

@aigenie Yes, I get it. I'm still struggling with this problem too. I even manage to generate 480p video in two steps with relatively good quality on the standard Wan2.1 GUI. But the dynamics of movement suffers greatly. I once caught a glimpse at the discussions of Lora FusionX, it seems like there is a node in comfyUI that normalizes CFG. We'll probably have to try to sort something out in this direction. Have you tried to work with Wan 2.2? But not yet

@TeosKuzen yes i am talking about with 2.2

@aigenie Oh, got it. I can't check 2.2 yet. But do the old Loras work in style? For some reason, I thought that Lora, which was usually installed on 2.1, would not clearly work on 2.2? Or is this not the case? I can give you approximate settings with a JSON example from the standard GUI for version 2.1 with more or less live dynamics. If necessary. I just don't know if it's suitable for 2.2. How will I start testing it - if I write anything!!! 

@aigenie Do you know any interesting loras to support video detail?

Were you already using another speed lora? This increases speed by reducing the number of required steps. Using speed lora can reduce motion and adherence

@Ada321 Thx a lot, man!

I get error This site can’t be reached I don't know whyyyyyyyyyyyyyyyyyyyyy !!!!!!!

30 minutes for 512x512 is too long I guess lightning loras is good.

after 1 hour I got wan 2.1 1.3b result thank you.

It allows you to generate Wan videos much faster and with fewer steps needed.

@Ada321 also my lora looks like shit when i try 1280x720

Yeah, i agree. I feel like nerds really get a kick out of jargon and try to invent as much as possible. It's annoying. But in the case of "self forcing", the general idea is that the generation reinforces the prompt at every step across the generation. So it winds up becoming a more amped up version of the prompt at the end of the generation as opposed to only lightly adhering to the prompt. This can create more dynamic end products.

That's my understanding, anyhow.

@jazgalaxy581 That is the first time this naming has been understandable to me, thanks.

@pufferjacketeven475 "Wan21_I2V_14B_lightx2v_cfg_step_distill_lora_rank64_fixed.safetensors" <- The self forcing file... About everything and his mother is mentioned but no the actual self forcing 🤣

Yeah, it took me a while to figure out lightx2v and self-forcing are the same thing. Or are they? Still not really sure. Then there's "lightning" which is also lightx2v? Then there's Kijai's versions, And FusionX which has lightx2v or self-forcing integrated, and there's causvid and accvid...😣

@GayLizardSpy No. As far as I understand, Lightx2v and self forcing are not the same thing.

The "t2v" wherever you see it, means "text to video". Or i2v meaning image to video.

In the case of "lightning", it's a... ugh... quantized(?) version of the file that will speed up the rendering by using a more focused model.

as far as I understand.

@jazgalaxy581 naming things is hard, It more than likely makes sense in the context of machine learning.

@GayLizardSpy Exactly :)

I2V when?

@Unscathed7928 Lightx2v said it's soon, NO date announce tho