Today marks the big day of training for a dataset that took me over 2 years to create.
I started a channel on youtube called @Dadsweathertapes, where I meticulously archived weather reports, as well as commercials and news segments spanning between 1986 through 2005.
I've been trying to come up with an easy way to sample the frames from all the work I did, and came up with a random video sampler script (attached to this article!) that takes inputs from a folder full of video, and attempts to create a 30 and 60 second 'sampler' video, composed of 3-10 second segments from every input video, including audio.
From there, I was left with 30 and 60 second clips from various runs of the script, which I then stuck into a ComfyUI workflow (attached!) to extract frames from the video, roughly 1 per second of video, and then perform a LUT color correction pass, to try and fix up the washed-out colors, and maintain a stable baseline. You may or may not want to do this, and can bypass the node completely in the workflow. I used it because I wanted a kodachrome style in the final outputs.
At this point, I went through the resulting frames and removed anything that was not an image, or too confusing to tag with gpt4v. I also used the Auto Tagger feature on CivitAI to tag images that had few or no tags. I also did a few other things to help tag the frames - adding the year and 'tv', 'commercial', and 'tv commercial' to each set of frames.
I've found that 3600-4000 steps with FLUX training results in a perfect style LoRA, for practically any input data. It's probably overtrained or over-fit, but the results that I've seen so far are very impressive at 1.0 or lower strength.
I separated each era due to the use of multiple VCR's - the first VCR was a Canon deck from 1986, purchased a month or two before I was born. The second was a JVC SVHS deck, purchased either in July 1992 or '93. The third was some no-name expensive SVHS deck that I wish I still had.
We have the following eras of VHS models:
Prompting guide:
Prompt for the following to turn on the LoRA's influence: tv, tv commercial, commercial
This goes for all the retroVHS models.
Years can help dictate a specific year, since the dataset was tagged with the appropriate year.