MiniMax H3 360° Orbit — Freeze & Motion

MiniMax H3 360° Orbit — Freeze & Motion

One image, two approaches to a camera orbit. This ComfyUI workflow uses the same prepared image as both the first and last frame, with a simple switch for a frozen subject or prompted subject movement. It combines Pablo Dawson’s 360° Orbit LoRA with LightX2V’s Turbo v1.2 LoRA.

Added MiniMax_H3_360_Orbit_Freeze_and_Motion_02_Auto_Midpoint.zip, containing the back-view guide workflow and sample front/back images. The back-view guide is automatically placed at the midpoint to help preserve rear details, such as shirt graphics and jersey numbers.

HOW TO USE

1. Load your image and select the required model files.

2. Set your seed, duration and maximum image dimension.

3. Choose Enable Subject Motion:

• false: requests a frozen scene while the camera orbits.

• true: uses the external “Motion Prompt - Subject Motion” text. Describe the action you want, such as dancing, gestures or a turn.

4. Queue the workflow and review the result.

Shared camera and background instructions stay inside the workflow in both modes: a continuous, single-direction 360° orbit, steady camera distance and height, and a background that remains fixed in world space. The external prompt controls subject action; you do not need to repeat the camera instructions. These are generation instructions, not guarantees.

IMAGE PREPARATION AND SETTINGS

The image is resized proportionally with Lanczos to a longest side of 768 px, then center-cropped so both dimensions are multiples of 32. This adds no padding or black borders, but may trim a small amount from the edges. Both resized dimensions must be at least 32 px; extremely thin images are unsupported.

Sampling settings and current duration:

• 6 sampling steps, CFG 1, Euler / simple, denoise 1

• Sigma shifts: 6 / 3

• Orbit LoRA: 1.0; Turbo v1.2 LoRA: 1.0

• Current Seconds setting: 4, mapped to 90 frames on the model’s frame grid

• Video output at that setting: 24 fps, approximately 3.75 seconds

The duration has been returned to 4 after a 5-second setting produced two revolutions in one observed result. This is a provisional choice; orbit count and subject motion still need further testing. Check the Seconds field after loading the workflow, as an earlier JSON revision stores 5.

“4step” is part of the Turbo filename; this workflow is saved with 6 steps.

The graph uses ComfyUI core nodes, including Math Expression and the If/Else switch, plus a packaged subgraph. No third-party custom-node dependency is declared in the JSON. Use a recent ComfyUI build with these nodes and MiniMax H3 support; older builds may show missing nodes.

MODELS AND CREDITS

Place files beneath ComfyUI/models/ in the following folders:

• diffusion_models/minimax/

DasiwaMinimaxH3_dasiwaHybridV2_int8.safetensors

Dasiwa reference: https://civitai.red/models/2877206/dasiwa-minimax-h3?modelVersionId=3374445

The linked page/version and its match to this filename have not been independently verified.

• text_encoders/

qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors

https://huggingface.co/Comfy-Org/MiniMax-H3/tree/main/text_encoders

• vae/

minimax_h3_video_vae_fp16.safetensors

https://huggingface.co/Comfy-Org/MiniMax-H3/tree/main/vae

• loras/minimax/

minimax_h3_flf2v_lora_v1.safetensors

360° Orbit LoRA by Pablo Dawson:

https://huggingface.co/pablodawson/MiniMax-H3-360-Orbit-LoRA

minimax_h3_fl2v_turbo_4step_v1.2_768p_comfyui_bf16.safetensors

Turbo LoRA by LightX2V:

https://huggingface.co/lightx2v/Minimax-h3-Turbo

Credit to the model and LoRA creators and Comfy-Org for the packaged components. This upload shares the workflow arrangement; the model and LoRA weights are the work of their respective creators. Please review the applicable licenses and usage terms on the model pages.

LIMITATIONS

The comparison clips use seed 42 with only Enable Subject Motion changed. Testing is limited. Exactly one complete 360° revolution, a stable background, consistent anatomy and seamless looping are not guaranteed, including at the provisional 4-second setting. Subject movement can introduce distortions or disrupt the orbit. Motion mode is an experimental extension of the frozen-orbit setup. The output is a generated video, with no 3D mesh export.