Creates videos in a stylized physically based rendering animation aesthetic, comparable to the kind seen in major animated films.
The primary trigger word for this style is "PBRSty1e," and there is no strict prompting pattern required; anything in natural language that conveys the concept "the final video is PBRSty1e" should be sufficient.
Out of the box it should support all input types and workflows available to the base model. If you are attempting to convert an existing video, the following boilerplate prompt can be used to do so:
subject_definitions:
<Video 1> is the source video containing the video that will be converted to PBRSty1e.
<Video Audio 1> is audio for <Video 1>
summary:
[reference generation] Convert <Video 1> to PBRSty1e. Keep the exact motion, timing, camera path, scene composition, subject descriptions, audio, and environment from <Video 1>.
retention_analysis:
<Video 1>: attribute_transfer for motion, posture sequence, camera, and environment only.
detailed_description:
[Shot 1] DESCRIPTION OF SHOT 1
[Shot 2] DESCRIPTION OF SHOT 2California AB 2013 Training Data Disclosure
This LoRA was fine-tuned using visual data consisting of both clips and still frames, which were AI-generated or extracted from publicly available clips and renderings. The training data includes copyrighted material owned by third parties, including both film studios and individual artists. No training data was licensed or purchased. This LoRA is provided for non-commercial use only under the terms of its distribution.
The dataset consists of approximately 2000 images and 50 video clips collected from publicly accessible sources and generated; dataset assembled in 2025-2026. Data was processed through standard resizing, cropping, normalization, frame extraction, and labeling steps. Synthetic images were included as part of the training dataset.
This model is intended for non-commercial, experimental, and educational use. Generated outputs may reflect copyrighted visual styles or themes associated with the underlying training data. Users are responsible for ensuring compliance with applicable copyright law, other intellectual property laws, and all other applicable laws.
Description
FAQ
Comments (4)
Leon turned into ParZival 😂 thanks ❤️
Would you be willing to share your captioning prompts for both images and video, as well as training settings?
Becuase no matter what i try to create a style lora for minimax it makes my motions slow-motion like, but your lora follows the video all till the end,
So any help would be very helpful, or tips and tricks :)
Thanks for your time anyways!
Of course! :) So this has a few different datasets, each with a unique approach:
- Set 1: Image pairs consisting of target style plus various other styles as Control1 (primarily photographs). The captions on these are just the style tag and nothing else
- Set 2: General tagged images, which are just like the target set, but don't have a source image associated with them. These are tagged as longform sentences in natural language that thematically describe every element EXCEPT the style (then I tack on the standalone style tag).
- Set 3: Videos aren't very important except in areas where the style overrides the underlying behavior, so it's actually a pretty small part of the datasets. Videos are tagged in the same way as Set 2 (natural language description of non-style contents + style tag). The most important element here is to ensure your clips are set to 24 FPS on disk (which you can fix with FFMPEG) and aren't so long that they'll need to be compressed at training time to fit in your VRAM (in other words, you want to be able to use "Auto Frame Count" in Ostris AI Toolkit). I also recommend that the clips be one continuous shot. You can normalize to 24 FPS and split by individual shots with the scripts I posted on HuggingFace for an article: https://huggingface.co/ReltivlyObjectv/video-dataset-processing-scripts/tree/main
@ReltivlyObjectv You are a legend! Thank you so much)