How to Turn an AI Video Glitch Into a Reusable Effect: The CardSpin Recipe

A small team caught MiniMax H3 misreading a Victorian portrait, and half an hour of training later the mistake worked on any photo. Here is the method, and what it takes to do it with your own happy accidents.

The most original thing an AI video model does is usually a mistake.

Anyone who generates video has a folder of them. The camera was meant to circle the actor and instead the whole room folded. The dog was meant to run and instead it melted into the grass in a way no animator would have thought of. You watch it twice, laugh, and delete it, because there is no way to ask for that again. Type the same prompt with a new seed and it is gone.

Yesterday a small team published a documented way to keep one. The release is called MiniMax-H3 ORB360 CardSpin, and it does something charming: it takes a portrait and spins it in space like a thin photo card, with the person inside turning with it. Profile on the edge, the back of their head on the back of the card, then a clean landing on the original photo.

The effect is fun. The story of how it was made is the part worth stealing.

The accident

The team, publishing as MATLOWAI, was building something ordinary: an add-on for the MiniMax H3 video model that makes the camera orbit a full 360 degrees around a still subject. They trained it on grey 3D renders from Blender. Then they tested it on real photographs, including an 1867 portrait of the astronomer Sir John Herschel by Julia Margaret Cameron.

The model got it wrong. Instead of orbiting a man, it decided the photograph was the object and turned the print itself in space, the man inside rotating with it.

Most people would have filed that as a bug. They filed it as a find. In their words on the model page, "It was too good to leave as a one-off."

What they did next

The method is short, and every step matters.

They took that one generated clip. Not a dataset, not a set of hand-made examples. One clip that the model itself had produced by mistake.

They wrote an honest description of what happens in it, with times. The published caption marks the key moments: the card goes edge-on at 1.0 seconds, shows its back from 1.7 seconds, reaches the other profile at 3.8 seconds, and lands back on the photo at 5.125 seconds.

Then they trained their existing orbit add-on for 50 more steps on that single clip. The card reports about half an hour on one graphics card.

It worked on photos the model had never seen: other Cameron portraits, and a colour photo of a cat in a hat. They also tried 100 and 150 steps. Those worked too, with slightly crisper card edges, but they released the 50-step version because it "has the most charm" and disturbs the original orbit behaviour least. The same file still does a normal orbit if you give it the orbit prompt.

Why one clip was enough

This is the part that makes the recipe usable, and it is easy to misread.

The half hour was not teaching the model to spin cards from nothing. The team had already spent 750 steps teaching an add-on to move the camera in a full circle around a frozen subject. That add-on already understood "rotate this 360 degrees and come back to the start." The glitch clip only had to teach a narrower lesson: the thing that rotates is the photograph, not the camera.

Think of it like teaching a dancer who already knows the turn. You do not start from walking. You show them one variation, once, and they have it.

The timestamped caption matters just as much. A vague caption like "a photo spins" tells the model nothing about pace. A caption that says what happens at each second gives it a timeline to hold on to. That is also why the effect lands on the original frame so reliably: the caption says it does, and when.

So the honest version of the recipe is this. One lucky clip plus a precise, timed description plus a short top-up, on top of an add-on that already does something close.

Put this into practice

There are two ways in. Use the released effect, or run the recipe on your own glitch.

Use CardSpin as it ships

1. In ComfyUI, load the MiniMax H3 model in its reference-image mode (the model page calls it Ref2VA).
2. Add minimax_h3_orb360_cardspin_step50.safetensors with the standard Load LoRA node at strength 1.0, applied to the model only.
3. Give it one portrait as the reference image. People and animals work best, and old sepia prints work as well as colour photos.
4. Paste the text from prompts/cardspin_caption.txt in the repository as your prompt.
5. Keep the clip at 124 frames and 24 frames per second. The prompt's timings assume that length, so a shorter clip will cut the spin off.
6. Roll a few seeds. Because it learned from one example, some seeds commit to the card harder than others.

One detail in their prompts is worth copying into all your H3 work. The two sound sections of the prompt are set to "N/A". The team notes that without that, the model tends to invent a soundtrack. If you want silence, say so.

If you would rather not install anything, the model page shows a hosted option through WaveSpeed, a third-party service. Expect to spend a few runs finding a good seed.

Run the recipe on your own glitch

This is where the real value is. The steps generalise to any effect you stumble into.

1. Keep the glitch. Save the clip, the seed, the reference image and the exact settings. You will want to describe it accurately later.
2. Find the closest add-on you already have. The accident needs a foundation. If your glitch came from a camera-move add-on, start from that one. If it came from the base model with no add-on, my guess is the top-up will need more than 50 steps and give less reliable results; the team did not test that case.
3. Write a timed caption. Watch the clip with a timecode on screen. Write what happens and when, in plain sentences. Name the subject. Say where it starts and where it ends. This caption is your instruction to the model, so write it the way you would brief an animator.
4. Train a short top-up. The team used musubi-tuner, a free training tool, together with the H3 training adapter from Ostris. Save copies at a few points (they compared 50, 100 and 150 steps).
5. Test on pictures it has never seen. An effect that only works on the original image has memorised a clip. It has not learned a move. Try at least three unrelated photos.
6. Pick by eye, not by the lowest number. The team chose the earliest version because it looked best and damaged the original behaviour least. Check that the add-on still does its original job too.

Where this breaks

Training needs serious hardware. The top-up run on this release peaked at 73 GB of graphics memory on a 96 GB workstation card. Using the finished effect is a much smaller job. Making your own means renting a large cloud machine by the hour, so budget for it.

One example means one way of doing it. The card always turns the same direction at roughly the same speed. The team is plain about this: it was "evaluated by eye on a handful of images and seeds, not with a benchmark." If you want variations, you need more examples.

The timing is locked. Change the clip length and the caption's timeline no longer fits, and the spin gets cut off or drags.

The foundation was not free. The orbit add-on underneath took 750 steps on rendered 3D scenes, at about 18 seconds a step. If you have no close add-on to start from, the "half an hour" figure does not apply to you.

The licence follows the model. This is an add-on for MiniMax H3, and it inherits the MiniMax H3 Community License, including its commercial and territory terms. The authors say plainly: "It is not MIT." Read the full terms before you use it on paid work, and apply the same care to any effect you train yourself on H3.

The habit worth building

The lesson here is not about spinning photos. It is about what you do with the folder of mistakes.

For most of us, AI video has been a slot machine: pull, look, pull again. The accidents feel like noise because they cannot be repeated. This release shows they can be captured, described, and turned into a tool. You need one clip, a description with times in it, and something close enough to build on.

That puts a new job in the creative workflow. When a model surprises you, the useful response is not "neat, next seed." It is to save the seed, write down exactly what happened and when, and ask whether it is worth a half-hour top-up. Some of the best effects of the next year will probably come from people who kept their mistakes.

Start the folder today. Next time the model does something no one asked for, keep it.

---

Medium metadata

Title: How to Turn an AI Video Glitch Into a Reusable Effect: The CardSpin Recipe

Subtitle: One accidental MiniMax H3 clip, a caption with timestamps, and half an hour of training turned a mistake into an effect that works on any portrait. Here is the method.

Tags: AI Video, Video Effects, ComfyUI, Generative AI, Creative Workflow

Estimated read time: 7 minutes