LoRAs don't seem to be working with the new Mega model
I've been trying out the Mega model, using the provided workflow, but I'm having major issues with getting LoRAs to work.
Wan 2.2 Low Noise Loras seem to get almost completely ignored on I2V, and even though they register on T2V generations, they're hugely unstable and provide very odd results. Wan 2.1 LoRas fare no better, working far less than they have in previous versions.
I've tried running them at 0.4, 0.65, 0.8, 1.0 and 1.5 strength, but nothing is working...
For the sake of testing, I even tried running Wan 2.2 Low and High LoRAs together - using 1.0 for the High and 0.5 for the Low - to see if injecting a bit of the high noise content without running at full strength might stabilise the LoRA a bit, but predictably, that worked no better than just running the Low noise LoRA...
Like I said, it almost seems like in many cases, the LoRA is being completely ignored...
the loras don't seem to be working on the GGUF version for me
I'm running the full version from this repo, so whatever is causing it seems to be affecting both the GGUF and non-GGUF versions
I don't think there are any "mega" GGUFs yet. I have not experimented much with other WAN 2.2 "low noise" LORAs yet, as "mega" is very very new. I would hope that WAN 2.2 low noise T2V loras should still work. Do not even try I2V "low noise" loras IF there is a T2V version (since the "mega" is based off of WAN 2.2 T2V "low noise").
I'm playing with the NSFW merge now, and the "mega" version seems to be taking my typical mix of NSFW LORAs....
Hmm, I'm using the regular non-NSFW version - the first one that was uploaded. I'll keep trying with it, but it just doesn't seem to want to play with any of the LoRAs, including ones that used to work pretty well with it.
Is this an NSFW merge? I just tried switching off the LoRAs and running a NSFW prompt without them and the result was a lot more NSFW than I was expecting 🤣
The "mega" non-NSFW version has no NSFW LORAs added at all, so if you are getting spicy content, it was already included in WAN 2.2 + SkyReels + FunReward etc.
I haven't posted a NSFW version yet, still messing with a LORA mix.
I can't wait for the NSFW version!
I was looking at a couple of WAN 2.2 LoRAs on Civitai just now, and they quite often seem to state that both the high and low noise LoRAs are needed as the LoRA won't work properly without both being active. Is there something within the AIO checkpoint make-up that overcomes this?
You know what, I have no idea what changed but I reloaded the workflow and tried a few LoRAs and not only do they seem to be working a lot better now, the video quality is outstanding visually...
I tried using the WanVideoWrapper with Kijai's fp8 scaled models and it destroyed my PC - when I did manage to get a result, it was low-res, ugly and just absolute trash.
I've tried GGUFs galore and had a similar experience.
But this - this is the first checkpoint I've used for Wan 2.2 where I actually feel like when I set 720x720, that's what I get back.
And I'm running on a 12gb 5070... yet a checkpoint that should be way too big for me to run performs an order of magnitude better than any GGUF or smaller FP8 model...
That's it, all those unnecessary diffusion models are going in the bin!
All to say... nice work Phr00t!
You know what, I have no idea what changed but I reloaded the workflow and tried a few LoRAs and not only do they seem to be working a lot better now, the video quality is outstanding visually...
I tried using the WanVideoWrapper with Kijai's fp8 scaled models and it destroyed my PC - when I did manage to get a result, it was low-res, ugly and just absolute trash.
I've tried GGUFs galore and had a similar experience.
But this - this is the first checkpoint I've used for Wan 2.2 where I actually feel like when I set 720x720, that's what I get back.And I'm running on a 12gb 5070... yet a checkpoint that should be way too big for me to run performs an order of magnitude better than any GGUF or smaller FP8 model...
That's it, all those unnecessary diffusion models are going in the bin!All to say... nice work Phr00t!
could you possibly share your workflow? the GGUF model are frustrating for me. i have an RTX 3090 24gb so i should hopefully get similair results to you.
Of course, let me tidy my workflow up a little bit, make it a little more SFW (🤣) and I'll upload it here for anyone who it might be useful to 👍
Okay, HF won't let me attach the workflow file directly, so I've included the PNG which should contain the metadata - if you drag and drop it into a blank Comfy window, it should load...
I've also attached a quick example of the video generated by the workflow as it is set up here - it took 162.87 seconds to generate a 5 second (81 frame) video at 640 x 640 👍
While T2V is working really well, I'm struggling with I2V - my prompts seem to be getting ignored 80% of the time and the results with and without LoRAs is... not great.
So if anyone has any sage advice or workflow ideas to crack this, I'd really appreciate it! 😊
the LoadLoraModelOnly is not loading lora´s for the Mega workflow. After taking a quick look in the console all keys was missing ... with the normal LoadLora node its seem to work. I get no Lora Key errors anymore.
edit: after more tests ... some lora´s seem to load and other dont.