Most AI video prompts produce a rectangle you watch. The MiniMax H3 360 equirectangular LoRA by rehan-fal aims to produce a sphere you can look around: front in the center, back at the joined left and right edges, sky and ground at the poles. It keeps H3's generated audio. This is a specialized projection adapter, not a new video model or a headset-ready file straight from the generator.
The adapter can make short monoscopic 360-degree video, and the creator provides sample clips, settings, and a packaging recipe. It does not make stereo depth or a walkable 3D scene. The separate Pepe104 Turbo Space uses a different LoRA and is not a demo of this 360 adapter. Most importantly, the base model's open-weight license currently excludes the EU, UK, US, and South Korea from its default territory; see the rights section before using or selling this workflow.
See the Creator's 360 Samples
For a fair comparison, inspect the same-prompt base-versus-LoRA image, then check the seam behind the viewer in an actual 360 player.
Credit: The adapter, measurements, and sample media are published by rehan-fal; the underlying model is MiniMax H3. The separate Turbo Space is by Pepe104. No test render or independent quality benchmark was produced for this article.
What the LoRA Actually Changes
An equirectangular image stores the entire viewing sphere in one flat 2:1 frame. Horizontal position maps around the horizon; the far left and right must meet behind the viewer. The top and bottom compress toward the poles. Rehan-fal trained a low-rank adapter on 181 screened 360 clips so MiniMax H3 is more likely to follow that layout instead of making an ordinary ultrawide shot. The underlying H3 model still generates the moving scene and synchronized audio.
The model card reports six-prompt geometry measurements: edge-seam correlation rose from 0.55 for base H3 to 0.95 for the new adapter, while pole spread fell from 0.89 to 0.26. Those metrics test projection consistency, not whether a scene looks realistic, is comfortable in a headset, or works for every prompt. They are the creator's small-sample measurements, not an independent benchmark.
What it is not: this is monoscopic video. Turning your head reveals other directions, but moving your head sideways does not reveal new perspective or stereo depth. The model card points to a separate VR180 adapter for forward-facing stereo content; VR180 and full-sphere mono solve different jobs.
A Usable 360-Video Workflow
- Clear the rights first. Check the H3 license and MiniMax's FAQ for your territory and use case. A public model page or Space is not itself a commercial authorization.
- Use the correct 360 route. The model card documents the
minimax/h3/text-to-video/lorahosted endpoint, including a direct adapter-file URL, and links a Space that explicitly lists this adapter. Check the provider's current access, billing, and terms before rendering. - Write for every direction. The creator's recipe starts with
eqr360and an equirectangular layout instruction, then describes what is ahead, left, right, behind, and overhead. The dedicated Space prepends the trigger and layout sentence for you. For example: “At the center of a fictional observatory, a telescope ahead, glowing control panels on both sides, a doorway behind, and stars overhead.” - Preview before 4K. The card recommends a 21:9 canvas, LoRA strength around
0.75, and a short five-second test. It says lower-resolution previews can help judge composition before a higher-resolution hosted render. Its stated cost and timing are a snapshot, not a current price quote. - Package and inspect. The published recipe scales the generated 21:9 video to exact 2:1, sets square pixels, and adds Spherical V2 metadata with Google's Spatial Media tools. Check the rear seam, horizon, poles, audio, and viewer comfort in a headset or 360-aware player. A flat MP4 preview alone is not the final test.
For a concrete starting configuration and the exact JSON payload, use the author's fal example. It recommends a direct weight-file URL rather than relying on the repo ID, because the author observed different outputs with the same seed when the two loading paths were used.
Where the Demo Is Strong, and Where It Can Fail
| Question | Evidence from the model card | What to test yourself |
|---|---|---|
| Do the edges meet? | Reported seam correlation is 0.95 with the adapter versus 0.55 without it. | Watch the point directly behind you throughout the clip, not just one frame. |
| Is the whole sphere covered? | Lower reported pole spread, and 4K samples show overhead and side views. | Look up and down for stretched objects and a misplaced horizon. |
| Does it hold over time? | The card says five seconds is tightest; longer clips may darken or drift. | Review each second and trim weak tails rather than assuming continuity. |
| Is it VR-ready? | The author supplies 2:1/Spherical V2 packaging steps and sample files. | Open the final file in your target 360 player and check sound and orientation. |
Its limits are unusually explicit: a faint rear seam can remain, close-up people may look odd because of first-person training footage, and no stereo depth is produced. The card says the sample files have seam softening, so raw renders may look less polished than the published examples.
The Two Spaces in Your Links Are Not the Same Demo
The 360 equirectangular Space lists rehan-fal/minimax-h3-360-equirect-lora in its model metadata. Its README says it prepends the 360 trigger, generates on a 21:9 canvas, and can package the output for 360 players. That is the most relevant interactive companion to the adapter.
The user-suggested Pepe104 MiniMax H3 Turbo Space is worth mentioning as a separate H3 experiment. Its README describes a split, unquantized H3 deployment and Turbo LoRA weights from larryvrh for fewer sampling steps. It does not list or load rehan-fal's 360 adapter. “UNCENSORED” is part of the Space URL, not evidence that the base model's license, rights, or use rules disappear.
The License and Training-Data Caveat
This matters particularly for readers in Portugal and elsewhere in the EU. The current MiniMax H3 Community License defines its default territory as worldwide excluding the European Union, United Kingdom, United States, and South Korea. MiniMax's FAQ says organizations in those places may apply for an authorized open-weight license; it separately says MiniMax's hosted API is globally available with safeguards. A third-party Space or hosted provider should not be assumed to have the same terms as MiniMax's own API. Verify the route you plan to use, especially for client work.
The adapter's data notice says the training set was selected from YouTube 360 videos, only a minority of candidates carried CC-BY licenses, and the original clips are not redistributed. The creator's position is that the LoRA learns layout rather than source content. That is a claim about training behavior, not blanket clearance for every output or commercial use. Use fictional or properly licensed subjects, disclose synthetic previews, and seek legal review when rights matter.
A Business Idea Built Around the Workflow
The interesting opportunity is not “sell AI VR.” It is a small, testable preview product for a buyer who needs to communicate a space or experience before it exists. Use this prompt to work from a reachable buyer, explicit rights, and a measurable pilot.
Test a synthetic 360-preview offer
Three buyer-specific options, then a seven-day pilot.
Source Map
- Adapter and creator samples: rehan-fal model card, sample files, and its geometry report.
- Relevant demo and packaging: 360 adapter Space and Spatial Media tools.
- Base model and rights: MiniMax H3 model card, community license, and territory FAQ.
- Separate Turbo experiment: Pepe104 Space, its implementation notes, and the different Turbo adapter.