shamanic/minimax-h3-equi360-lora

MiniMax H3 · Equirectangular 360° LoRA

A LoRA for MiniMax H3 that generates full-sphere, monoscopic equirectangular video with native audio. The trigger is equirect360. Describe the environment, motion and sound after the layout prompt.

Reviewed v2 · September 2026

The default weights now use the reviewed 57-clip dataset, covering 36 independent sources and 16 scene families. This replaces the cave-heavy v1 checkpoint previously published here.

The dataset was rebuilt with stricter watermark/overlay filtering and checks for disconnected scene content both at the wrap boundary and inside the frame. Diversity was retained across live action, game/CG, anime and generative imagery. Minor upscaling seam artifacts were retained where the scene content connects; polar distortion and those upscaling artifacts were tagged in the affected captions. Vocals were retained.

We compared this run against a separate 119-clip candidate using four matched prompts and seeds: coast, cabin, synthetic garden and an anime character. Both looked good in flat and inside-sphere views; the reviewed run was marginally preferred. This is a small qualitative comparison, not a benchmark. The 119-clip candidate is distinct from the earlier published v1.

Checkpoint	File
Current default: reviewed v2, step 2500	h3-equi360-lora-step2500.safetensors
Versioned copy of the same reviewed weights	h3-equi360-reviewed-v2-step2500.safetensors
Previous published v1	versions/v1/h3-equi360-lora-step2500.safetensors

Reviewed v2 SHA-256: a1e3ebd2b79be92d5e970c6f516519f8bd7f8e4271a71e7a7d7c8e938e157cf7.

Samples

Scene	Generated panorama	View from inside the sphere
Coast	Video	Camera tour
Cabin	Video	Camera tour
Synthetic garden	Video	Camera tour
Anime character	Video	Camera tour

The tours loop each short generated video at normal speed while the camera turns through 360°. Exact prompts, seeds and generation settings are in training/reviewed-v2.json. Older files directly under samples/ are v1 examples.

Using the model

Start at LoRA strength 1.0, with the H3 FL2VA base. The evaluated generation settings are 21:9, 768P, 5 seconds, with prompt expansion disabled.

{
  "prompt": "equirect360 Equirectangular 360-degree panoramic video: the frame is a full spherical projection, a complete environment wrapping around the viewer on all sides, with the horizon at the vertical middle of the frame. A quiet cherry-blossom garden with a wooden bridge and a pond, stationary camera. Birds sing and leaves rustle.",
  "loras": [{"path": "https://huggingface.co/shamanic/minimax-h3-equi360-lora/resolve/main/h3-equi360-reviewed-v2-step2500.safetensors", "scale": 1.0}],
  "aspect_ratio": "21:9",
  "resolution": "768P",
  "duration": 5,
  "prompt_expansion_mode": "disabled"
}

This is the input used with fal's minimax/h3/text-to-video/lora endpoint. The complete 21:9 frame represents the full sphere: resize to 2:1 when packaging it as conventional equirectangular media, and add mono-360 spherical metadata for players that require it. The sample camera tours map the full frame directly onto the sphere.

Training

Setting	Reviewed v2
Trainer	fal minimax/h3/t2v/trainer
Steps / rank / learning rate	2500 / 32 / 0.0002
Training bucket	Medium, 21:9, 124 frames at 24 fps
Adapter	208 BF16 tensors; attention qkv_proj and out_proj in 50 transformer blocks and 2 token-refiner blocks
Dataset	57 clips, 36 sources, 16 scene families; includes 7 recovered archive clips
Render domains	21 generative, 15 live-action, 14 game/CG, 7 anime
Caption tags	19 clips tagged for polar distortion; 6 for upscaling seam artifacts
Final reported training loss	0.5134

The raw training media are not redistributed here. The repository's older pipeline scripts and v1 model card describe the earlier work; the released v2 checkpoint and sample settings are recorded in training/reviewed-v2.json.

Limitations and licence

Wrap discontinuities, polar stretching, unintended motion and other generation artifacts can still occur. Dataset filtering does not guarantee watermark-free outputs. Quality tags describe the training data; they are not a demonstrated negative-conditioning control. Inspect the result in a spherical view for its intended use.

The weights derive from MiniMax H3 and remain subject to the MiniMax Community License.