Stop AI Video From Duplicating Props and Pushing Hands Through Tables: Count-Everything Prompts Plus the VBVR Add-on for MiniMax H3

Watch enough AI video and you start counting apples. A customer puts three in a bag, and somehow there are four on the counter afterwards. A hand reaches for a remote and passes straight through the coffee table. Someone chops a pepper, then chops it again in the wrong order. The image looks like a film. The physics looks like a dream.

These failures kill more takes than ugly lighting ever does, because they are the ones a viewer notices without knowing why. And they are hard to fix after the fact. You can grade a bad colour. You cannot easily grade away a fourth apple.

This week a small studio called Licon Studio released an add-on for MiniMax H3, the free video-with-sound model that a lot of ComfyUI users (ComfyUI is the free node-based app many people run AI models in) have adopted since the summer, aimed squarely at this problem. It is called Licon VBVR, and its model card says it improves object consistency by "reducing unintended object duplication, clipping, and interpenetration." That is a useful promise. But the part of the release I keep coming back to is not the add-on file. It is the six test prompts the author shipped beside it, because they show a way of writing video prompts that helps whether or not you install anything.

What the add-on is

Licon VBVR is a small file (149 MB, per ComfyUI Wiki) that plugs into H3's image-to-video mode (you give it a starting frame and a description, it makes the shot). Add-ons like this, technically called LoRAs, nudge a big model toward a particular behaviour without replacing it. Their "strength" setting is a dial from off (0) to full effect (1.0).

The author trained it on a rebuilt version of VBVR, a collection of video clips built to teach models how objects and actions should behave: things stay single, stay solid and happen in a sensible order. According to ComfyUI Wiki's write-up, the training set listed 434,894 items at 1280×720, and the sound part of H3 was left untouched, so your audio comes out exactly as the base model makes it.

Using it takes no special software. Drop H3-Licon-VBVR-I2V.safetensors into ComfyUI's models/loras/ folder and load it with the standard LoraLoaderModelOnly node inside whatever H3 image-to-video workflow you already run.

The add-on itself is tagged Apache 2.0, a permissive licence. H3 is not. More on that below, because it decides whether you can use any of this on paid work.

The prompts are the method

The repo includes a validation folder with six test shots. Each has a starting image, the original H3 render, the render with the add-on, and the prompt. Read the prompts and a pattern jumps out. They are written less like descriptions and more like a continuity supervisor's notes.

Take the fruit stall. The prompt sets the stage with "exactly three red apples on the counter." It describes the customer opening a paper bag with the left hand and placing each apple with the right, one at a time. Then it adds the rules: "The vendor does not add or move any fruit," and "Keep the count consistent throughout; no cuts, duplicates, or extra apples." Even the sound design counts: three separate apple thumps.

The living-room test does the same thing with a magazine and a blue remote. One person, one magazine, one remote. The magazine goes to the lower shelf of the coffee table first, then the remote goes beside the TV. And then the fence: the person "does not turn on the television, sit down, or move any other object," with "no cuts or added objects," and "all background objects remain unchanged."

That structure has four parts, and each one shuts a door the model likes to walk through:

1. An inventory with numbers. Not "some apples," but exactly three. Models are loose with quantity unless you pin it.
2. Actions in order, with hands. Step one, then step two, left hand, right hand. A numbered sequence leaves less room for the model to reshuffle events.
3. A "does not" list. Name the things that should stay still and the actions that should not happen. The model cannot avoid a mistake it has no reason to think about.
4. Sound tied to action. One sound per event ties the audio to the picture's count, which matters on a model that makes both at once.

None of this needs the add-on. It is just good direction. The add-on is the second layer: a model nudged toward solid, single objects, given a prompt that has already told it what solid and single mean in this shot.

Put this into practice

Here is the routine. It works best on the kind of shot that usually fails you: hands handling objects, two people passing something, anything with a countable prop.

1. Pick one shot you have lost takes on. Pour a drink, plate a dish, hand over a package. Make a clean starting frame with the props visible and countable.
2. Write the inventory. List every prop that matters with a number. "One white mug, one glass bottle, one folded napkin."
3. Write the actions as a numbered sequence. Say which hand, and what happens first. Keep it to two or three actions for a short clip.
4. Write the fence. One sentence on what does not change ("nothing else on the table moves"), one on what does not happen ("she does not drink from the mug"), and "no cuts or added objects."
5. Write the sound to match. One sound per action, and say "no music" if you want it clean for the edit.
6. Render without the add-on, with a fixed seed. The seed is the number that controls the randomness. Keeping it fixed means any difference in the next render comes from the change you made, not luck.
7. Add the VBVR add-on at strength 1.0 and render again with the same seed and prompt. Scrub both side by side at the moments of contact: the grab, the set-down, the handover.
8. Try lower strengths. The author gives no guidance on strength, so test 0.6 and 0.8 as well. If 0.6 holds the props as well as 1.0, use the lighter setting; it is less likely to change things you liked.
9. Keep a note of what changed. Fixed the count? Fixed the hand clipping? Changed the actor's face? A three-line note per test saves you rediscovering it next month.

Before you download anything, open the validation clips in the repo and watch the before and after pairs. Ten minutes there tells you whether the improvement is the kind you need.

Where it breaks

The evidence is thin. Six test pairs, all chosen by the author, all everyday photo-real scenes like the fruit stall, a kitchen and a living room. There are no measurements, no failure cases and no limitations section on the card. Treat it as promising, not proven.

Nobody has mapped the strength setting. The card offers no guidance, so your tests are the documentation for now.

It will not touch sound. The sound part of the model was not trained, so if H3 makes a fourth apple thump, the add-on will not save you. Your sound-tied-to-action prompt is the only lever there.

It only covers image-to-video. The file is trained on H3's start-from-a-frame mode. Making a shot from text alone, or from a set of reference pictures, is outside what the author tested.

You still need H3's hardware. The add-on is small, but H3 is not. If H3 already runs on your machine, this will too. If it does not, the add-on changes nothing.

And the licence is the real wall. The add-on is Apache 2.0, but you run it on MiniMax H3, and H3's licence grants use only outside what it calls Excluded Territories: the European Union, the United Kingdom, South Korea and the United States. People in those places have to contact MiniMax for a separate licence, and commercial use above US$20 million a year in revenue needs MiniMax's written permission. If you are in one of those regions, this is a learning exercise until you have that paperwork.

The bigger point

AI video has trained a lot of us to reroll. A shot comes back with an extra cup, so you hit generate again and hope. That habit treats the model as a slot machine.

The VBVR release points somewhere better. The add-on is a sign that the community is now fixing specific, nameable failures in open models within weeks of launch. The prompts are a sign that the fix starts with us: write the shot the way a script supervisor would track it, count what is on the table, say what must not move, then test one change at a time.

So pick your worst shot. The one where the props never behave. Rewrite its prompt as an inventory, a sequence and a fence, render it on a fixed seed, then add the VBVR file and render again. Whichever version keeps the count, you will have learned something about this model that no launch reel would have told you, and you will spend less of next week counting apples.

---

• Sources: LiconStudio/Minimax-H3-VBVR · VBVR validation prompts · ComfyUI Wiki on the VBVR add-on · MiniMax-H3 licence