Why the base model freezes, and why that is the opportunity
First–last-frame generation exists to pin down both ends of a clip. Reference-to-video treats input images as loose appearance references, so a clip can end anywhere. FL2VA fixes the end frame in advance, which is what you want for loops and for stitching shots together.
It creates an obvious edge case. If the first and last frames are the same image, the simplest reading of the job is that nothing happens, and that is roughly what the base model does: the creator's published comparisons show it barely moving. Give it only the first frame and the opposite happens: the camera moves, drifts, and never comes back to the starting view.
The LoRA sits in that gap. It is trained to give identical end frames one specific interpretation: the scene is frozen, and the camera is what moves, all the way around, back to the start. Whether it delivers that on your photo is a separate question, and the orbit is the creator's claim, not ours.
*Testing status:* This guide documents the configuration published by the LoRA's creator. alfw.ai has not downloaded, installed or run the adapter, and has not independently reproduced the demonstration results. Settings, compatibility and performance claims below are attributed to the creator unless otherwise stated.
The workflow, step by step
*1. Load the FL2VA base, not the reference variant.* The card specifies MiniMax-H3 FL2VA, pruned, using the Comfy-Org int8 ConvRot repack, with the LoRA applied at strength 1.0. The card also says the LoRA was trained and tested with ostris/ai-toolkit's MiniMax H3 extension, and that its keys use ComfyUI naming. Treat that as the creator's reported setup, not a compatibility guarantee: before running it in ComfyUI, confirm that your installed MiniMax H3 nodes support this adapter's weight naming and loading format. The creator's training environment does not by itself establish that every ComfyUI workflow can load the adapter.
*2. Set the same photo as the first and the last keyframe.* This is the step that feels wrong and is the entire point. Identical end frames are the instruction the LoRA was trained to read as an orbit.
*3. Use the author's prompt, word for word.* The card says "use this prompt verbatim," and the prompt is written to suppress everything except camera movement:
"One frozen instant. Only the camera moves. In a continuous 360 orbit. Preserve every person and object in exactly the same world position, orientation, shape and pose throughout the shot. Airborne objects remain suspended at the captured height and angle: no wobbling, shaking, spinning, drifting, falling or continued action. Keep faces, hands, clothing, liquids and the background motionless while retaining their natural appearance. Camera parallax is the only source of apparent movement. No cuts, zoom, morphing or added objects."
*4. Use the author's recommended settings, at least for the first run.* 768×768 resolution, 73 frames — about 3.04 seconds of encoded video at 24 fps, with 72 frame intervals (3.00 seconds) between the first and last images — 28 steps, audio off, and no guidance: H3 is guidance-distilled, so there is no CFG or negative prompt to add. Changing several of these at once makes a failed result hard to read. The published recipe is the baseline; departures from it are your experiment, not the author's claim.
*5. Check the return, not just the orbit.* The adapter's stated job is a final frame that matches the first closely enough to loop or stitch. Identical end frames are conditioning inputs, not a guarantee: they do not by themselves establish a geometrically consistent orbit or a visually seamless loop, and the motion in between can still drift. A clip that orbits beautifully and lands a few degrees off has given you a nice shot and missed the job it was built for.
What the 28 clips actually show
The training data is the most instructive part of the card. The creator did not scrape orbit videos. He collected hand-selected human Gaussian splats — static 3D captures of people — and rendered an orbit video around each one. The renders are generated from 3D representations with consistent camera trajectories, although rendering artefacts and reconstruction errors may remain. What that construction buys is a dataset in which the camera is the thing that changes: no acting, no physics and no background movement inside the lesson.
The creator trained the adapter on those 28 orbit renders. The published examples suggest the LoRA learned that pattern within a narrow domain — short, square, human-centred shots. No broad independent evaluation establishes how reliably it transfers to arbitrary photographs.
The training run itself was modest by current standards: rank 16, alpha 16, applied across 208 transformer modules; 3,000 steps at batch size 1 — roughly 107 passes over the 28 clips, an approximation that assumes one distinct example per step, not a measured epoch count; AdamW 8-bit at a learning rate of 1e-4; one NVIDIA A100 80GB for 10 hours 24 minutes. That A100 figure is the creator's training hardware, not an inference requirement: the card does not state a minimum GPU for running the adapter, and nothing here establishes that it will fit a consumer card. A separate training adapter was used during training only and is not needed to run the LoRA.
The method is still worth copying even if you never use this particular file: build training data in which the one thing you want to teach is the only thing that changes.
Where it will probably disappoint you
The creator lists the limitations plainly, which is rarer than it should be, and they deserve to be taken at face value.
The domain is narrow. Twenty-eight human-centric, square, 73-frame clips define what has been demonstrated. A building, a car, a pet, a widescreen frame or a ten-second orbit are unverified applications — not proven failures, and not demonstrated successes either; nothing on the card settles them. The second warning is smaller but matters for the frozen-time premise: subtle movements and blinks may still appear, which will break the effect in exactly the shots where a viewer's eye is resting on a face.
There is also a limit no setting removes. A single photograph does not show the back of anything. The views an orbit passes through behind the subject are inferred by the model, not recovered from the photo: plausible generated geometry, not a reconstruction of what was actually there. For a stylised reveal that distinction may not matter. For a product shot that has to be accurate, it does.
And three separate licence questions sit underneath this tutorial; they should not be merged. First, the underlying MiniMax H3 weights are governed by MiniMax's Community License, which restricts use by territory — a point alfw examined when the weights shipped, in the licence piece linked in the sources below. Second, the model card reviewed for this article did not clearly state a licence for the LoRA itself: do not assume the adapter is unrestricted merely because it is publicly downloadable. Third, the practical step that follows from both: check the applicable base-model licence, including any geographic restrictions, before local or commercial use. Hosted access to H3 is a separate question with separate terms.
The practical version
Treat this as a specialised adapter with an unusually honest manual. Load FL2VA, duplicate your photo into both keyframe slots, paste the prompt unchanged, run the published settings, and judge the clip by its landing frame. If you need that orbit for a person in a square frame at about three seconds, the creator's documentation is currently a short, reproducible starting point from a still to a loopable shot. If you need anything else, you are no longer following a tutorial. You are running the next experiment, and the card gives you the training recipe to run it properly.
Sources
- [1] pablodawson, "MiniMax-H3 360° Orbit LoRA (first–last frame)", Hugging Face model card — all settings, the verbatim prompt, training details and limitations in this piece are from this card, read in full 19 October 2026Read source
- [2] Demonstration reel (discovery source for this piece), Instagram, posted 19 October 2026 — four subjects rotated from a single photo each, linking the same Hugging Face repoRead source
- [3] Background, alfw.ai's earlier coverage of the base model's licence terms: "MiniMax Called H3 Open. The License Says Otherwise in Four Regions.", 3 October 2026Read source