Session: ROMP-NEAT

Expected run time

140 minutes

Instructions:

  1. Don't use upscaled video.
  2. Separate multiple URLs with a space.
  3. A mix of 4K and 1080p resolution is preferred as it significantly enhances the model's definition and range.
  4. The video URL can link to YouTube videos, channels, or playlists, and is compatible with various other video hosting platforms.
  5. The video source should be at least five minutes long and your subject should be the main focus.
  6. Note that the model's fidelity is highly dependent on the video's clarity, lighting, and recording angles.
  7. For YouTube content, we can download and process up to 10 videos from channels and playlists, including YouTube Shorts.
  8. Select the base model you would like to train a LoRA for.
  9. The base model selection will impact runtime, price, and available options.
  10. Pro Tip: Select "Data set only" to make sure the training data is what you want before training the LoRA.
  11. Upload media directly to One Shot.
  12. Re-use a data set you've trained with us before.
  13. If you have downloaded a dataset from us in the past, it will show up here and you can use it to train a new LoRA.
  14. Enter a distinctive name to serve as the trigger that activates your LORA model.
  15. This name acts as a unique identifier, allowing you to summon your model with a specific word or phrase in any prompt.
  16. You can upload individual media files or ZIP archives containing media files.
  17. A mixture of 1080p and 4K media is recommended.
  18. Upload a MINIMUM of 25 images or 25 seconds of video.
  19. If we do not detect at least 20 faces, the job will fail. Faces must be at least 300 pixels tall or wide to be detected.
  20. No matter how much you upload, we will only extract a maximum of 500 frames for processing. 8.5 minutes of video or longer will yield 500 frames.
  21. Check this box to save the dataset we train your LoRA with.
  22. You can re-use a saved dataset later by selecting it under the "Show more options" section.
  23. Generate 4 sample images using your new LoRA with custom generated prompts based on the imagery found in the training data.
  24. Face detection is enabled by default. It is the magic behind our character LoRA process.
  25. Leave face detection enabled unless you are trying to make a style or concept LoRA.
  26. Face detection will remove images that do not contain your subject.
  27. Face detection will crop face and body shots from all images that match and that are big enough (at least 300px wide or tall).
  28. Lower values mean more similarity and fewer matches. Below 20 nearly all shots of the same person will not match.
  29. Higher values mean more tolerance and more matches. Above 80 most different people will match each other.
  30. If your subject's face is the most frequently shown, we can select it for you automatically.
  31. Even if your subject is not the most frequently shown in any one video, as long as it is more common than any other face across all source videos for this job, we will find it.
  32. To specify a specific face, upload a reference image.
  33. This option makes the dataset NSFW prior to training.
  34. Your original images will be digitally altered to remove the clothing from the subject in the images.
  35. A LoRA trained on this NSFW dataset will generate high quality NSFW images more consistently.
  36. The dataset can be fun to look at, too.