One LoRA, two camera tricks for MiniMax-H3 Ref2VA. Pick the mode with the prompt.
Card spin Trigger: ORB360_CARDSPIN
Your photo turns in space like a thin photo card: profile on the edge, the back of the head on the back, then it lands back on your photo. Uses one photo.
360° orbit Trigger: ORB360_CW
A smooth clockwise camera orbit around your frozen subject, ending on the starting view. The photo's surroundings are kept. It is fairly easy to prompt for a grey studio background as well for subject isolation.
Works from one photo, or better from three views of the same subject (for example three renders or a turnaround sheet): the front, the view 120° clockwise around it, and the view 240° around. With real side and back views the model does not have to invent them.
How to use
- Load LoRA at strength 1.0 on the MiniMax-H3 Ref2VA model.
- Paste the matching prompt as the text (all on Hugging Face, link below):
- card spin, one photo: cardspin_caption.txt
- orbit, one photo: orbit_realscene_caption.txt
- orbit, three views: orbit_3ref_caption.txt
- In ComfyUI, use the MiniMax H3 Reference to Video node. Connect the photos in order: ref_image_1 is the start view (Picture 1), ref_image_2 the 120° view (Picture 2), ref_image_3 the 240° view (Picture 3). Keep ref_image_size on "match".
- 124 frames at 24 fps; the prompts' timings assume that length.
- The prompts keep the output silent. Roll a few seeds.
Prompts, example files and the full training notes:
https://huggingface.co/MATLOWAI/MiniMax-H3-ORB360-CardSpin
The card effect was learned from one clip made from a public-domain 1867 photograph; this is not a likeness model.
Licence: MiniMax H3 Community License Agreement. Powered by MiniMax H3.
Description
Retrained at a higher resolution. The card effect now reduces sharpness so if you need 360s to be sharp use the 1500 step version.
FAQ
Comments (3)
Amazing! Thanks so much for posting this. I can confirm that version 2 works even better - probably gives perfect results 90% of the time compared to maybe 50% for version 1.
Grok seems to be rather good at doing 360s. Maybe something for your training data? (NSFW) https://civitai.red/posts/30016442
Thank you! That's amazing! I look forward to your release of versions with both high and low angle perspectives. I'm currently using your LoRa to generate videos to assist in creating Gaussian models in a project, and while we're still optimizing it, it's already showing great promise.