MiniMax H3 ORB360 and CardSpin are community LoRAs for two different shots: a camera circling a frozen subject, and a flat photograph turning in space. Their shared repository makes it easy to download the wrong file. Start by choosing the motion you need, then choose the adapter and reference recipe that describe it.
This guide uses MATLOWAI's model card and its published prompt files, reviewed on October 13, 2026. The instructions below explain the documented setup; H3Vid has not run this adapter or benchmarked its results. Our online generators do not currently offer ORB360 or CardSpin as presets.
Choose the file before changing the prompt
The repository contains three approximately 597 MB, rank-32 adapter files. They require MiniMax H3 Ref2VA base weights; downloading an adapter alone does not provide a video model.
| Desired shot | File | Important distinction |
|---|---|---|
| Clockwise camera orbit around a frozen subject | minimax_h3_orb360_step1500.safetensors | The author's preferred orbit checkpoint |
| A photograph rotating like a thin card | minimax_h3_orb360_cardspin_v2_step50.safetensors | Adds card motion training to the step-1500 orbit checkpoint |
| Reproduce the original card effect | minimax_h3_orb360_cardspin_step50.safetensors | Uses the older orbit checkpoint as its starting point |
The CardSpin variants can respond to orbit prompts, but the author reports softer results in a controlled cat-photo comparison. That does not establish a universal quality ranking: image-detail statistics from one clip cannot predict every subject or seed. It does establish a practical starting choice. Use the orbit file for an orbit, and the CardSpin file when the photograph itself should turn.
One image is supported; three views add information
A single reference image works with the published real-scene orbit prompt. The camera has to invent surfaces hidden behind the subject. Three reference views give it additional evidence: the front, a view 120 degrees clockwise from the front, and a view 240 degrees clockwise.
This matters when reading community comparisons. ORB360 is not restricted to three-image input. Conversely, supplying three pictures does not turn a generated clip into a measured 3D reconstruction. Choose views of the same rigid subject, with consistent pose, materials and lighting. Three unrelated character portraits are not three camera positions around one frozen scene.
Preserve reference order. The three-reference caption assigns the opening view to the first picture and the intermediate orbit milestones to the next two. If a workflow rearranges the files, its text and images no longer agree. Inspect the actual conditioning connections before queueing.
Documented ComfyUI setup
Begin with a working MiniMax H3 Ref2VA workflow and its required base components. This adapter is not an FL2VA checkpoint and should not replace one silently. For the base environment, see our MiniMax H3 ComfyUI overview.
- Place the selected adapter where your ComfyUI installation loads LoRAs. Record the full filename so the original and v2 CardSpin files remain distinguishable.
- Use the standard Load LoRA node on the model branch. The model card specifies model strength 1.0 and describes compatibility with ComfyUI's H3 weight mapping.
- Connect the prepared reference image or ordered views to the Ref2VA workflow. Keep the subject frozen rather than requesting a turn of its body.
- Choose the matching file from the repository's
promptsdirectory: the real-scene orbit, three-reference orbit, or card-spin caption. - Retain the documented timing for the first reproduction: 124 frames at 24 fps. The shown cat example uses 1024×768, 20 inference steps and seed 20260926.
- Save the graph, prepared inputs, prompt, adapter filename and generation settings with the output. These are documentation-based instructions, not a workflow we have tested.
The captions include separate subject, summary, retention, shot and audio sections. Both audio sections request silence. Keep them when reproducing the source recipe; removing them introduces another variable and may allow an invented soundtrack. Requesting silence in a prompt is different from disabling an audio execution branch.
Frame milestones explain the timing
The three-view prompt anchors its references as follows:
| Zero-based frame | Time at 24 fps | Camera angle | Reference |
|---|---|---|---|
| 0 | 0 seconds | 0 degrees | First picture |
| 41 | 1.708333 seconds | 120 degrees | Second picture |
| 82 | 3.416667 seconds | 240 degrees | Third picture |
| 123 | 5.125 seconds | 360 degrees | First picture again |
The last timestamp describes frame 123, not an arbitrary duration slider setting. A 124-frame export at 24 fps has a playback duration of about 5.167 seconds. Changing frame count or playback rate while retaining these milestones changes the recipe. Use the original caption first, then adjust times, angles and view labels together when designing another shot.
Camera height has a narrower validation boundary
The author tested held-out Blender objects against known rendered orbits. The reported tests use three objects and one seed per case. Some changes in camera elevation worked on grey-backdrop renders; large height jumps and elevation changes during the clip failed.
Real photographs have a separate limitation. In the author's tests, requests to orbit above or below an eye-level photograph stayed near the photograph's original height. Do not treat a successful rendered-object example as proof of reliable high-angle control on real scenes. Start at the input's camera height and review any requested offset as an experiment.
Review the journey, not only the seam
Scrub through the back view and both intermediate references. Check that the camera travels clockwise without reversing or inserting cuts. Watch whether the subject remains rigid, whether the background changes perspective continuously, and whether hidden surfaces stay consistent after their first appearance. Returning to the first image can hide a shortcut through the middle.
For CardSpin, use a different acceptance check: the photographed content should remain on a visibly thin rotating surface. If the scene instead becomes volumetric, that may be an interesting effect but it missed the requested card behavior.
The pablodawson 360 Orbit LoRA is a separate FL2VA adapter with matching first and last images and a different documented sampling recipe. Neither adapter's name proves pixel-perfect loops or industrial geometric accuracy. Our Keyframes and Motion Control pages explain related controls, but do not imply these community weights are selectable online.
The adapters carry the upstream MiniMax H3 Community License, including its territorial and use provisions. Check the repository's complete terms before choosing a deployment. Your next practical step is to select one shot, download its matching adapter and caption, and reproduce that fixed configuration before introducing extra LoRAs or acceleration methods.
The cover is an AI-generated concept illustration of reference views and a spinning photograph. It is not an ORB360 result, a CardSpin result, or a ComfyUI screenshot.