For on-site generation, I recommend only using models designed for WAN2.1 or earlier. It does not appear that Civitai can get 2.2 models to work. You can use the Universal model with both high and low noise steps, but if you want the best quality, I included both versions as well.
This LoRA is trained on conversational language. You can describe your scene in plain English or Chinese, and it should adhere to your prompt well. Below are some inspirations for you if you need it.
Positions:
On her back
Kneeling
Perspectives:
POV
Enhancement phrases:
She bounces her body up and down
She bounces her breasts up and down with her hands
All showcase videos have the workflow metadata embedded. You can download the video and drag it into ComfyUI to replicate them with their prompts, seeds, and settings.
Description
Optimized Training for I2V
FAQ
Comments (11)
Even for I2V, you should train on the T2V model. It has the benefit of working on T2V and I2V. A lora trained on I2V only works on I2V.
This would be why I release an I2V specific trained model, and will release a T2V version when it's done. There's a significant difference in quality between the 2 for their respective usages.
I included a comparison that shows how significant the movement difference is.
Also, you can use the I2V model with T2V, with less accuracy.
@ComfyTinker considering that the previously posted titfuck Lora looks much better than yours even on i2v and it was trained on t2v...I would agree to disagree with you. There is absolutely no use in training on the I2V model.
@switch_ai "Looks much better than yours even on i2v" is a strong statement with no evidence to support.
The 2 example videos are 0-seed examples, meaning there was no prompt-fishing to find them. Additionally, I even provided a side-by-side comparison of the 2 with the same seeds and it's clear the i2v model has better adherence.
@ComfyTinker you showed an example of I2V vs I2V optimized... And I agree that "optimized" is much better but that has nothing to do with what I said.
@switch_ai The example is the original V1 model run on I2V vs the newly published I2V optimized model.
It's literally a comparison of the "V1 WAN14B" model versus the one that I just published.
@ComfyTinker I get that. And what I'm saying is that if your use the T2V model for training it has the benefit of working with T2V and I2V but if your train with the I2V model, it only works with I2V and NOT T2V.
@switch_ai And I'm telling you that I'm currently training a T2V model that will be published later. If you want to only download the T2V model, that's fine. However, the I2V-trained model works 10x better than the T2V model, which is why I trained it.
@ComfyTinker confidently incorrect
@switch_ai Liking your own messages doesn't make your point look any better. You clearly do not understand the dimensional difference between I2V LoRAs and T2V LoRAs. It's quite ridiculous.
My man. Enjoying your PK lora, this one looks great too. And thanks for the time you spent making your tool guide
Details
Files
Titfuck_WAN14B_V3.0__I2V-Optimized_by-ComfyTinker.safetensors
Mirrors
1592364_Titfuck_WAN14B_V3.0__I2V-Optimized_by-ComfyTinker.safetensors
Titfuck_WAN14B_V3.0__I2V-Optimized_by-ComfyTinker.safetensors
Titfuck_WAN14B_V3.0__I2V-Optimized_by-ComfyTinker.safetensors
Titfuck_WAN14B_V3.0__I2V-Optimized_by-ComfyTinker.safetensors
Titfuck_WAN14B_I2V_by-ComfyTinker.safetensors
Titfuck_WAN14B_V3.0__I2V-Optimized_by-ComfyTinker.safetensors
Titfuck_WAN14B_I2V_by-ComfyTinker.safetensors
Titfuck_WAN14B_I2V_by-ComfyTinker.safetensors
Titfuck_WAN14B_I2V_by-ComfyTinker.safetensors
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.