Interesting. So this is sort of like a lora(?) that is intended to lower the amount of steps needed to produce a quality result like the Wan lightning lora? https://huggingface.co/lightx2v/Wan2.2-Lightning/tree/main

I wonder its stance on NSFW material

@Cyph3r how did you tested it? comfyui is not loading the model.

@Cyph3r In my tests, all of my videos were better without it, using lightx2v on its own.

@Cyph3r I will test your workflow. Before it gave me several errors of the lora not loading on the log.

@Cyph3r So, it does not work with normal nodes, should be used a kijai setup wrapper only, or can it be used with Power Lora Loader?

@Cyph3r do you have a workflow you can post? the loras are not getting loaded in my usual workflow

its uses custom scheduler - https://github.com/Yaofang-Liu/Pusa-VidGen/blob/main/PusaV1/diffsynth/schedulers/flow_match_pusa_v2v.py

+ info from: https://github.com/Yaofang-Liu/Pusa-VidGen/blob/main/PusaV1/README.md

https://huggingface.co/Kijai/WanVideo_comfy/tree/main/Pusa

wan 2.2 versions are less than 1 gb each.

tyvm, but i got the lora weigh error using the 900mb one

@fronyax It doesn't work bro. Wait for updates.

It worked with kijai wanwrapper nodes for me. They also updated nodes for pusa yesterday but that update caused my 5090 to go 100% and got stuck there repeatedly. So I went back to previous version. It works but I don't think worths the hustle. It supposedly do I2V with T2V model but in terms of consistency I2V is still much superior imo.

same. i tried the lora here and kijais rescaled lora. absolutely zero difference

alright just saw that comfy isn't loading the loras at all. they might not work with the fp8/gguf models:

lora key not loaded: blocks.0.cross_attn.k.alpha

lora key not loaded: blocks.0.cross_attn.k.lora_A.weight

@crombobular You're right. I just realized that I have the same error messages (I'm using GGUF models)

I can't get this lora to load either. Tried both T2V and I2V. Not using GGUF.

It doesnt work lora key not loaded: blocks.0.cross_attn.k.alpha

lora key not loaded: blocks.0.cross_attn.k.lora_A.weight

@crombobular I downloaded the files from the huggingface page and it loaded. I did not get any errors. Differences were subtle and I preferred the generation without it 100% of the time.

@Cyph3r The sample workflow is a Wan 2.1 I2V workflow and there are no nodes in there that I did not already have installed... So what is the custom node you are referring to? What is it called?

@playtime_ai typical that there is still no reply on how to magically make this work

@Fealow Why are you directing this comment at me? This isn't my model and I didn't post it.

"their huggingface sure doesn't explain anything"
yea - it does not, but https://github.com/Yaofang-Liu/Pusa-VidGen/blob/main/PusaV1/README.md does

They claim it improves quality but I didn't see it in my testing. I preferred every generation without it over the one with it.

From my own experience, it "enhances" motion. Videos that don't have much movement will have more movement if you enable Pusa. This can be good, or bad, depending on the type of video you are generating, but generally it helps.

@Mu5hr00moO it actually doesn't explain anything in that link either, its just showing light2x and talks about a custom node that includes a noise slider, which by itself DOES improve quality at higher steps, at the cost of distorting the first frame, but you don't need this lora to do that.

"In Da Pussa!!"

Giant Mexican Laugh erupts.

What does CineScale do??

@fronyax I'm not entirely sure, but somehow it boosts the quality — mainly the resolution.

@fronyax i used it few times, it was mostly related to a movie like cinematic scene, where the camera work looks more dynamic, it also bring a different visual tone to the videos.