Numina teaches AI video models how to count

Ask any modern AI video generator to put two kids, one dog, and a cat into a scene, and you will almost certainly end up with three kids and zero cats. Even top-tier models like VO 3.1 and Grok Imagine constantly fail basic arithmetic.

Researchers just dropped the code and paper for Numina, a training-free framework designed to force text-to-video diffusion models to actually respect the exact object counts in your prompts.

Fixing the layout without retraining

Generating precise object counts has long been a headache for generative video. Usually, fixing this requires expensive fine-tuning or throwing massive datasets into post-training, which most smaller teams cannot afford.

Numina takes an identify-then-guide approach instead. Basically, it peeks into the diffusion process, picks out the most discriminative self- and cross-attention heads, and extracts a rough countable layout from the latent space.

If that layout does not match the prompt—say, showing four cyclists instead of three goats—Numina conservatively refines the positions and tweaks the cross-attention layers to steer regeneration.

On the new CountBench benchmark, this lightweight intervention bumped counting accuracy on Wan2.1-1.3B by up to 7.4%. On the larger 5B and 14B models, it scored gains of 4.9% and 5.5%, all while keeping temporal consistency intact and boosting overall CLIP text alignment.

Plug-and-play for open source

The framework is model-agnostic, meaning it can hook into virtually any open-source diffusion backbone. The authors have already published working implementation guides for Wan2.1 and Wan2.2, complete with setup instructions for running the code locally.

If you already run video diffusion models on your own hardware, you can plug Numina right on top of the inference pipeline without retraining a single weight.

Why it matters

For real-world video production, count inaccuracies are dealbreakers. If a client script calls for two protagonists walking through a door, a clip showing three people is instant junk.

We think Numina's approach is brilliant because it tackles generation errors at the inference stage. Instead of burning millions of dollars in compute trying to teach a model basic math from scratch, it acts like a lightweight guardrail during generation.

A 5% to 7% boost in accuracy will not magically solve every multi-subject hallucination, especially when fast motion or heavy occlusions come into play. But it turns blind seed-rolling into a structured, controllable process, which is exactly what open-source video tooling needs right now.