For WAN 2.2 inference.
If you enjoy this pack and want to support my work, you can buy me a coffee here:
👉 buymeacoffee.com/designedbycrt
This is my RUNPOD referal link to earn 5$, wich is about 6h of free RTX 5090 usage : https://runpod.io?ref=u7b2habt
It also helps me to train new models, thanks.
Description
This Pack Contains:
INGREDIENTS for specialized features:
14 High-Noise Models
10 Low-Noise Models
7 Face Models (HuggingFace link included)
I recommend euler 8/12 steps, shift 5/8.
When using multiple LoRA in a stack, try to balance the strenghts to not overcook your renders
FAQ
Comments (36)
Am I reading that correct, if we use the "stacked" LORA then we don't need to also use the lightning LORAs because they are baked in?
If you talk about the merged models, SSH and SSL, it's exact, to save VRAM and compiling time before inference
@pgc Thanks, and what is IL style referring to?
@BinaryBottleBake IL style was trained on 1,500 suggestive "anime" style images (100% synthetic), but only for the high-noise model. As a result, the final video remains realistic while featuring more stylized, with fantasy-like proportions.
"SS_SHIRT_NO_NIP" can be used to prevent nipples bulging on bra/shirts etc on the low noise inference.
excellent work! works great with I2V too,
Thanks, I don't think using it with i2v is a good idea, since it was trained on t2v with the corresponding sigmas, i2v uses different settings and training scripts, but I didn't test so I can't really tell
@pgc well it works pretty well, i mean.. you can tell its for T2V but still good and most of your pack doesnt change the face which is always nice :)
@sirbazzrick I trained an I2V model, mostly for motion (skin bending, physics, etc..), But I still don't know where I should upload it, I will see
Brilliant Models been using these a lot with my T2V creations with the SmoothMix wan Models and they are fantastic, good job hope to see more models from you.
my comfy ui disconnects when i use merged lora, works fine with single train version. I am using Q4 K M high and low with lightning 4steps lora, I have tried using merged lora with and without lightning 4 steps 4, it always disconnects. am i missing something? please help
If your comfyUI disconnect it means that you reached the maxiumum RAM available, the merged models are fp8 safetensors and may be too big for your config.
The merged models already have the last lightning models, so I don't recommend adding it afterwards with a lora loader.
Free more disk space and use larger swap file
You mentioned in the description about a huggingface link to the 7 models you included. Did you have a link? Just wanted to see what those were. thanks!
In the zip.
Any gguf version?
this is crazy
This is marked as being a lora. Is that true? This feels like a checkpoint to me.
H and L checkpoint merges (dyno+scaled_fp8), bunch of loras with which those checpoints were merged, and one all-in-one lora. Literally i wrote the same what was in description.
@forfreelsd368 Ah, okay. I thought at first that they were just listing an ingredients list, like I've seen elsewhere. I see now that it's actually in the zip.
I guess it's a difference in perspective, regarding labeling. With the size of it, it feels like the "main" thing being shared here is the checkpoint, since it's all in one zip and can't be avoided. That makes the loras feel like side-notes rather than the main thing being shared. But yeah, I guess there's no clear label to use here, since the loras are indeed valuable on their own.
@Jellai well, I'd say that those loras even valuable for those merged models, because merges sometimes not up to the task.
Doesn't tried loras all the way or on other models, damn I even don't know where to get a time to test THIS models. -_-
@forfreelsd368 Frankly, I chose not to release every LoRA I initially trained. Each model requires comparisons, proper descriptions, and further testing, which would have significantly increased the size of the pack. I was concerned that a larger quantity of models might be perceived as a lack of quality and inconvenience those downloading the zip file.
The main content is the individual LoRAs, the merged models serve as a template to demonstrate how you can conserve VRAM and reduce compiling time during inference. They also provide a general idea of the strength to use when stacking multiple LoRAs to avoid undesirable results. I encourage you to experiment with your own stack and merge it to suit your unique preferences.
@pgc I thought it's because of civitai loves ban some stuff today. ^_^
Also heard, that while merging - some stuff, like layers, will dissapear, so merge - isn't just model+loras even if loras at strength 1. So I was sad and dropped that.
In recent weeks, I've been trying to create some lora (but lack of local training guids for 16vram pisses me off), but I got distracted by creating another dataset, then another, and then all sorts of authors are releasing interesting models with loras (which, perhaps, I wanted to create), and somehow it goes like this all the time -_-.
I think it would be most effective to separate just the checkpoints (not as zip) and just the loras (zip possible). 21gb of the 28gb in the zip goes to what is essentially the output of a 6 node workflow.
@firemanbrakeneck agree; this is unwieldy in its current form.
@pgc Thanks for that.
What is the update? Says it was updated today?
It left early access
why i gotta wait this long for titties man, dayum
I sense a theme....
It would be a good idea to upload the high and low noise models separately. Most downloaders have trouble with zip files, and the model becomes inaccessible to site generators. I, and others, would appreciate this. I know you don't want to upload the versions, but it would make this much easier to access and use.
I tried running the loras, replacing lightx2v on both the high and low side with the high/low "single train" loras included in this zip. The results were extremely blurry. Is this intended to be run with a high step count?
The single train loras doesn't embedd lightning loras, only the merged models have lightning models in it
So some of the videos I see here are plasticky. Is that the initial images that have been poorly made that made it plasticky? or is it the video gen itself?