Likeness of Nicole Eggert, early 90's.3gg3rt to trigger.Edit: As requested, I've attached the training date. Note that some images blurred for obvious reasons. We'll see if this violates ToS. Technically it should not since they are blurred. I've also updated information about how I trained this model. Will add this for every lora i upload going forward.V2.0: I'm much happier with this model. I updated training based on @kmdcomp's advice to only use the keyword token for image captions. Results are much much more consistent. I also update my generation workflow. Really happy with the resulting generations!V1.0: Closeups seem to work well. Medium/far shots are hit and miss. Used JoyCaption2 for captioning this time (as opposed to manually captioning), which may be why likeness is harder to achieve. Had to play with prompts and multiple generations. Shelving this for now, so I thought I'd at least share with where it's at. Feedback welcome.<h3 id="training-sgb3bmxoo">Training</h3>I trained this lora on a 4090 using <a target="_blank" rel="ugc" href="https://github.com/tdrussell/diffusion-pipe">diffusion-pipe</a>. I followed <a target="_blank" rel="ugc" href="https://civarchive.com/articles/9798/training-a-lora-for-hunyuan-video-on-windows">this</a> tutorial to set diffusion-pipe up on my machine.Character/likeness loras seem to be best trained with images, while activities seem to best be trained with short videos. Then in order to mix the character with the activity, use both a character lora and an activity lora.I've experimented with various captioning approaches: joycaption2, manually captioning using simple descriptions, and trigger word only. Trigger word only has produced the best results by far since it seems avoid baking in any pre-conceived information that might cause the generation prompt to I've attached the training data (faces blurred) to give people an idea of the various compositions I used. The images I used are a mixture of full body/clothed, full body/not clothed, and close ups of her face. I used rope to perform face swaps. I think face swapping can be a super useful way to fill out a training data set, clothed or not-clothed, as long as you can find a model that has similar body type, facial structure, hair. Also, if you don't have sufficiently detailed face shots, use an image up scaler workflow to achieve the desired detail. Or if you're capturing screencaps, you can use davinci resolve with denoise + super-scaler to improve resolution.Training took around 8 hours.<h3 id="generation-smysubbve">Generation</h3>I've embedded workflows in the example videos. Download and drag a video to Comfyui. Use Comfyui manager to install missing nodes.Prompting: Use 3gg3rt as the keyword. Do not describe any features of her person/body (eg: don't use 'blonde straight hair', 'skinny', etc) since her likeness is already baked into the lora and using these terms often results in a different likeness being generated.For example:"3gg3rt standing in front of a pool. She is wearing a red t-shirt and cut off jeans shorts."Also, prompting for specific articles of clothing or body parts often helps frame the shot. If you want a full body shot, prompting for feet or shoes usually works. You can use 'full body shot' or 'close up shot', but I find that 'full body shot' doesn't always work.

Nicole Eggert

Files with SHA256: d31ce6434ea789e7e45dd95ff314f3e583e642f29b8e4e4c209f5c1c3249ff83