V2 -------
Trained on a bigger dataset. It is better at the "obscured camera" look but not as good at the changing motion. Added the trigger word CandidChange. Could still be good for stacking with other Loras for the "obscured camera" trick but I have not tested it much yet. For some things V1 may still be better but V2 should be better with front facing. I will try to make a merge of both if i can ever figure how to merge Hunyuan Loras in Comfy. Here is the random prompt I have tested with using RandomPrompts node:
CandidChange. A woman is {dressing|undressing|adjusting her clothes} in a {bathroom|locker room|bedroom}. She is {partially naked|wearing a bra and panties|wearing a casual outfit}. The camera view is {partially hidden by a stucture|filmed through a hole in the wall|taken from an open doorway|seen through a window|obscured by a foreground object}.V1 -------
Candid videos of dressing and undressing. There is no specific trigger, specific descriptions seem to work well. "A woman is getting dressed" or "A video of a woman changing her clothes" followed by descriptions of the video quality like "The camera is obscured" or "the video was taken from behind an obstacle."
Description
FAQ
Comments (22)
I'm getting distinct square censoring/blurring of the face in my generations. I'm assuming you blurred out faces as part of training, but it seems that has been learned and carried over into the final lora.
ethical creeping
Looking forward to version 2
I include in my captions that the person has their face censored or blurred. That's worked for me to keep it out of the lora. I've only trained Hunyuan char loras using images, but it seems like the same would apply to vids
I actually did not do any face blurring, so this is a curious result. Did you stack with other Loras?
@CZTehTron Very odd. I was using the magic wand lora, so that could have been the cause. Likely the prompting of "obscured" triggered it some.
@KinkMaster I agree, I think I need better training prompts. "Obscure" does not seem like the right word to use, it seems like it made problems for others as well.
@CZTehTron personally, I liked the censoring and am probably going to make a lora for doing exactly that lol. Love your lora either way. Use it in all sorts of forms.
@KinkMaster Yeah my Magic Wand LoRA was trained with blurring and it does start to pick it up if the training goes too long. I tried to minimize this by using some out of frame faces as well, of course captioning "her face has been blurred out" (If you want to reproduce the effect on purpose, that's how it's captioned and Hunyuan on it's own understands it even without a LoRA trained this way.), limiting LoRA rank etc. It's rare that I get a blurred face without asking for it, but it does happen(1/15 maybe) and yeah a word like obscure might be semantically associated enough to trigger it.
thanks for the Lora, please consider including full prompts of your examples somewhere.
Thank you, I'll try to include full prompts in future updates :)
Thanks for making this lora.
When I use "taken from behind an obstacle", it shows my subject from behind/facing away from the camera. Did you experience this?
Some of the training data is from behind, but this makes me think I need to train better prompts that are more clear about positions and things.
Can you try training this on WAN 2.1 I2V 480p? It seems to have better base knowledge of dressing.
Now that I think about it, this would be way better combined with WAN 2.1 T2V 14b. I hope you'll consider it at some point.
I haven't tried WAN yet, but might be fun to mess around with!
@CZTehTron Absolutely. At first I was not sure... now I can't go back because it's so much better (even though slower) than HV.
@CogDiffusion I will try it soon, I am very curious now! But I am not patient, I use the Fast model for HV (which is probably why my examples videos are not great hahaha)
@CZTehTron it's not fun lol
@CZTehTron Once you find the right parameters and use things like Sage-Attention and Torch compile and Teacache it's not that slow anymore! I recently upgraded Torch and my speed doubled... I still can't properly comprehend that. Absolute game changer for me. Just generated a video with LoRA at 240x432x77 (great res for prompt testing) with 1.84s/it.
@CogDiffusion What version of torch are you on? I'm on 2.6 for cuda 12.6. Python 3.12.8. Running Triton, Sage, Tea cache, and more on a 24gb 4090FE on Windows. Torch compile absolutely fucks my iteration speeds and causes tons of inference errors, so I don't use it for anything. Torch compile is inherently messy in how it works, and is not very great for installs on Windows OS.
@KinkMaster Sorry to hear that. I am no Torch 2.8 nightly now and with the Python 3.12 version of Comfy it works like magic. Made the error of changing the Kijai's Compile Node from "Default" though and got an error as well. Dynamic makes it slower for me as well. The default definitely helps a lot though.