VR-360-Outpaint 1.0 — LTX-2.3 IC-LoRA
VR-360-Outpaint started as a proof of concept built at ADOS Paris in April 2026 and was first published on 23 April 2026. Version 1.0 followed in early July 2026, released under a non-commercial license; commercial licenses for it are now available from Burgstall Labs. Development has continued since, with a new architecture and a substantially larger, higher-quality proprietary dataset.
Burgstall Labs is the sole authorized licensor of VR-360-Outpaint. Commercial use of 1.0 or any later version, including offering services built on it, requires a license from Burgstall Labs (licensing of the LTX-2 base model from Lightricks applies separately). Burgstall Labs also offers conversion to immersive formats, including VR360, VR180 and Domemaster: licensing@burgstall-labs.com
This model performs 360° equirectangular video outpainting. Give it a flat video crop projected onto a black equirectangular canvas (inverse gnomonic projection, done by my nodes), and it fills the unknown regions to produce a full 360° equirectangular video suitable for VR / immersive playback.
Built on Lightricks/LTX-2.3-22B via IC-LoRA (video-to-video conditioning). This is the production "v1.0" release.
How it works: the inverse gnomonic projection trick
VR-360-Outpaint is built on an inverse gnomonic projection of the input, an approach I first developed in April 2026. The flat video is projected onto the equirectangular canvas at its true angular position and field of view, so the model sees exactly where the known pixels sit on the sphere. Instead of having to work out the spherical mapping and the outpainting at the same time, the model only has to complete the rest of the sphere consistently with that window.
The ComfyUI-VR-Outpaint-Tools node pack performs this projection from your flat footage, with automatic FOV estimation via the GeoCalib node.
Quick start
- Base model:
Lightricks/LTX-2.3-22B(dev safetensors) - Load this LoRA on top via the LTX-2
video_to_videopipeline (ComfyUI recommended) - Trigger word:
equirectangular - Caption (the only one the model knows):
360 degree equirectangular panorama, high quality, detailed - Resolution: 1920×960, 121 frames @ 24 fps
- Workflow: Use the included
Burgstall-VR-Outpaint.jsonin ComfyUI - Companion tools: ComfyUI-VR-Outpaint-Tools provides the gnomonic projection nodes and the GeoCalib FOV estimation node for automatic field-of-view detection from your flat footage — no manual FOV entry needed.
Examples
Each example shows the flat input video, a slow gnomonic pan across the outpainted 360° result (a virtual camera sweeping through the scene), and the raw equirectangular output.
| Flat input | Pan across the result | Raw equirectangular output |
Which checkpoint?
Three checkpoints are provided (.safetensors, ~1.3 GB each). They all produce valid equirectangular
output; pick based on your priorities:
| Checkpoint | Step | Character |
|---|---|---|
360vroutpaint_v2_step07000.safetensors |
7000 | Recommended. Best balance of equirectangular consistency and scene variety |
360vroutpaint_v2_step05000.safetensors |
5000 | Slightly more creative outpainting, but the 360° wrap can have a more prominent seam |
360vroutpaint_v2_step09000.safetensors |
9000 | Tightest equirectangular consistency, but tends toward more uniform / monotonic scenery |
The differences are subtle; if unsure, start with step 7000.
Training data
VR-360-Outpaint 1.0 was trained on a private, curated dataset of 360° video. All source footage is used under licenses that permit this use (Creative Commons Attribution), and the license records are archived. The dataset is not distributed with these weights.
Creator credits for the source footage are listed in ATTRIBUTIONS.md.
Sweet spot
The model was tuned on and performs best with:
- Content: Semi-static establishing shots — cityscapes, urban scenes, landscapes
- FOV: ~90–110° horizontal
- Aspect ratio: 1.78–2.39 (standard widescreen to cinemascope)
- Camera motion: Slow pans, gentle dolly, static tripod
It will generalize beyond these conditions, but quality degrades gracefully the further you move from them.
Limitations
- Top & bottom caps (zenith / nadir of the sphere) are the hardest regions; the model has less signal there. Avoid critical detail near the poles.
- Busy motion (fast cuts, whip pans, handheld shake) easily break the wrap seam.
- Prompt adherence — conditioning is dominated by the reference video, not text. The base caption / prompt is good to keep there, but you can describe the elements that you want there to be in the outpainted area to guide it. It will outpaint with the base prompt only, too.
- Narrow domain — trained on outdoor / architectural scenes. Interiors, crowds, close-ups, and organic subjects might work less reliably, although these draw strongly from the model's base training.
Companion tools
- ComfyUI-VR-Outpaint-Tools —
A ComfyUI node pack providing:
- Gnomonic projection nodes for converting flat footage into the masked equirectangular reference this LoRA expects (FOV selection, projection, canvas compositing)
- GeoCalib FOV estimation node for automatic field-of-view detection from your footage — no manual measurement needed
- Formerly named ComfyUI-EquirectProjector; renamed after expanding scope beyond the original idea of just forward gnomonic projection.
Burgstall-VR-Outpaint.json(included in this repo) — End-to-end ComfyUI workflow wiring the tools → LTX-2.3 + this LoRA → output.
⚠️ Repo re-star request: The node pack was renamed from
ComfyUI-EquirectProjectortoComfyUI-VR-Outpaint-Tools, and in the process I accidentally turned the repo private for a moment — which resets the star count to zero. If you starred it before, I'd really appreciate it if you could drop by and re-star it if you appreciate this work. Every one helps! ⭐
See also
- ComfyUI-Seamless-Equirectangular — seamless 360° equirectangular image and video generation (seam and pole handling for LTX-2.3 and other models), also by Burgstall Labs.
Acknowledgements
- Compute sponsorship by Lightricks — the LTX team graciously provided the GPU compute for training VR-360-Outpaint 1.0. Special thanks to Ofir Bibi for technical guidance and advice throughout the project.
- Initial PoC training done at ADOS Paris 2026 with collaborators Cseti, NebSH, and S4f3ty_Marc. Thank you for your collaboration and support.
- IC-LoRA training methodologies and multi-reference dataset design advised by oumoumad.
- Built on the Lightricks LTX-2.3 foundation model.
License
This VR-Outpaint IC-LoRA is a Derivative of LTX-2 and is distributed under the LTX-2 Community License Agreement, with Burgstall's additional terms. Under those terms, non-commercial use of Burgstall's contribution is permitted. Commercial use of Burgstall's contribution requires a separate commercial license from Burgstall Labs.
Commercial licensing inquiries: licensing@burgstall-labs.com
Users are independently responsible for complying with all applicable Lightricks licensing requirements.