ComfyUI-Flat2VR
Nodes for turning a flat photo into a VR180 image with the flat2vr (MiniMax-H3-Image) and l2r (Qwen-Image 2.1) LoRAs.
ComfyUI-Flat2VR
Nodes for turning a flat photo into a VR180 image: the flat2vr LoRA (MiniMax-H3-Image) makes the left eye around the photo, the l2r LoRA (Qwen-Image 2.1) makes the right eye.
Install
cd ComfyUI/custom_nodes
git clone https://github.com/troysniper01/ComfyUI-Flat2VR
Or install it from ComfyUI Manager (search "Flat2VR"). Needs a recent ComfyUI (the workflow uses subgraphs). GPL-3.0.
Models:
- MiniMax-H3-Image:
minimax_h3_image_pruned_int8_convrot,qwen3vl_32b_minimax_h3_nvfp4_awq,minimax_h3_image_vae_fp16 - Qwen-Image-2.1:
qwen_image_2.1_int8_convrot,qwen3vl_8b_int8_convrot,qwen_image_2.1_vae_bf16 - the flat2vr and l2r LoRAs in
models/loras
Workflow
workflows/flat_to_180vr.json. Place the photo in Flat2VR Place Photo, optionally fill in Flat2VR Prompt, and run.
It goes Left Eye → Hi-res Fix (re-rendered at 4096, back to 2048) → Right Eye → Soften → Join Eyes → Vignette.
Output is 2048 per eye, side by side. If a VR player doesn't detect it, add _180_LR to the file name.
Sampling as trained: H3 6 steps, l2r 40 steps, both cfg 1, euler / simple.
Nodes
- Place Photo: puts the photo on the 180° canvas (drag to move, scroll to resize, Shift to crop).
- Prompt: builds the prompt in the training-caption format.
- Paste Photo, Refine Mask: put the original pixels back; mask for the hi-res pass.
- Right Eye (best of N): makes the right eye and retries seeds until it passes the stereo checks.
- Join Eyes: left | right, with brightness and divergence fixes.
- Soften: a slight blur that brings H3's sharpness down to the right eye's.
- Vignette: fades each eye's rim to black.
- 180 Preview: look around the result.