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
For WAN 2.2 I2V Inference
Trained on a large set of trimmed videos (5s), that will enhance the body physics on your I2V setup
FAQ
Comments (30)
Absolutely phenomenal. Thanks for all the work and time. Works great I2V and combined with other Loras!
A really great merge, ngl!
Having a lot of fun just throwing prompts from previous generations made with base model and different loras into your merge and see what it creates. Quality is amazing with 12 steps!
27gb file? what.
Did you even try to read the description? :)
@Lora_Addict yeah but 27gb
The high and low noise merged models take up most of that space, each are 13gb. It also comes with a bunch of loras that you can pick and choose, each enhancing different properties. I personally enjoyed experimenting with this very much.
@zerox369 The ingredient loras are all included in the merged models right?
@Lora_Addict yes, I think they are all included at a specific strength. The MODEL_MERGE folder has a picture with more details. I've been trying the individual loras with base Wan2.2 and they also work great there with added flexibility of choosing the strengths. Some of the best jiggle physics I've ever seen XD
@zerox369 Thanks, that picture is helpful! Yeah i really like this merge, having a lot of fun and so many ideas!! :D
Yes, it's a large zip file. CivitAI doesn't allow multiple files under the same version, which is why I packed them all together, it's more convenient for me. I don't want to deal with the mess that would come from creating a separate version for each individual LoRA (since not every high model necessarily has a low-noise equivalent, and vice versa), as that would cause even more confusion.
@zerox369 The merged models don’t have all the individual “ingredients” baked in. I generated a large number of videos with different settings, but the merge I provided is somewhat soft to keep it stable and more general-purpose
@pgc thank you for the clarification and wonderful models. truly good stuff
This is great, can you share the workflows in the gallery so that we can follow along? I can tell that they were altered in Topaz Video or something to do interpolation due to the file name suffix, but can't copy it over to ComfyUI. A version of the videos with the metadata preserved would help bunches. Thanks for your efforts!
Hi. What does the I2V Skinbend Lora do? I don't see it mentioned in the description. I assume it adds a more realistic body elasticity and movement. Thanks!
I didn't make a proper description yet, but it adds some body features with chest / glutes physics, ( blowjob / handjob / breastfuck ) are also part of the dataset but I didn't test these features yet, but it's there
@pgc awesome, looking forward to trying it out.
所以，它对于生成示例图中的胸部类型有辅助功能吗，我很喜欢这个胸部效果
@pushpibl It's an I2V lora, you need to generate your own images prior with an text-to-image based model.
I only tested one of the sample videos but it didn't contain a workflow. Anyone have weight suggestions for the merge?
You can load SSH or SSL into comfyUI with drag and drop to load the model baking workflow with the current weights and lora models used
@pgc Possible misunderstanding, I meant the weights to use when using the 0_train or merged lora.
All loras works at the range of 0.15 / 1.5 strength, for a stack, it's better to balance them rather than using 1 on every models. But there is no universal values to set
@pgc You keep mentioning stacks, I thought this was a merge of different LORAs. There'd only be the one weight for the low and one weight for the high...
Whats with the Vram usage? i downlaoded your secret sauce pack and opened up your t2v from your workflows. best i can do is 368x480 81 frrames on my 4090 mobile , on smoothmix t2v im doing 720x1024 and thats at 121 frames :/
am i supposed to be ussing SSH and SSL?
@seanhan19911990198 You can use both SSH and SSL, or use the base models with your own stack configuration, SSH and SSL baked models use the same amount of vram than any other wan 2.2 models.
For information, a 4090 desktop can barely render 640x1024 for 81frames with 15 block swaps, so I have doubts about 720x1024 121 frames on a 16gb GPU, but maybe you talk about the TI2V 5b model, wich is not the model used to train these LoRAs
@pgc https://ibb.co/sdTtMtDx have a look mate. it will take around 8-10 mins to do this image.
67%███████████████████████████████████████████████████████▎ | 2/3 [05:37<02:56, 176.48s/it]
this is without any block swap or any speedup
@seanhan19911990198 Tbh I can't do this resolution on my end (24gb), maybe for 81 frames with some more block swaps, but 121 is too much. Even with the base models, so I don't really know.
As you can see when you drag and drop the smoothmix model (low noise on this example)
The merge is regular, it contains vae and clip https://ibb.co/4w1vG9Fd
On mine, vae and clip are not embedded in the .safetensors output, this is why mine is 13.3go,
and the smoothmix one is 19.8go.
But there is not much difference except for the base model used. I use the dyno model (high noise) from lightx2v:
https://huggingface.co/Kijai/WanVideo_comfy_fp8_scaled/blob/main/T2V/Wan2_2-T2V-A14B-HIGH_4_steps-250928-dyno-lightx2v_fp8_e4m3fn_scaled_KJ.safetensors
And this low noise model, with the last T2V lora from lightx2v:
https://huggingface.co/Kijai/WanVideo_comfy_fp8_scaled/blob/main/T2V/Wan2_2-T2V-A14B-LOW_fp8_e4m3fn_scaled_KJ.safetensors
This does not gives you an answer, but I can't tell you much since the merge SSH and SSL performs identically as the default models on my machine
@pgc sorry for the confusion but are you saying i could be using the smoothmix as a checkpoint instead of a diffusion model? im not super familier with everything ive been using a seporate clip and vae
@seanhan19911990198 yes since both vae and clip is inside already, it should work with a checkpoint loader
you really shouldnt be using 121 frames on a model trained on 81 frames 💀