flat2vr (MiniMax-H3-Image)
The node pack is required to use this properly.
Turns a normal photo into the left eye of a VR180 image. Place a photo on a half-equirectangular canvas, and the LoRA fills the rest of the 180° view around it. It was trained on one fixed instruction, so use it as is via the Flat2VR Prompt node. Use it with the 180 to Stereo 180 LoRA in the full workflow for a 180 VR image.
Training
A dataset of 3,800 pairs at 2048x2048 from real VR180 photos.
Rank 384, lr 1e-4, batch size 4, 2048 px, loss masked.
The data leans NSFW and has position trigger words.
Limits
Anime and cartoons don't work very well.
Level or slightly downward placements (pitch −15° to +5°) work best.
Fine detail can look harsh or crunchy because of H3's poor VAE.


