Nvidia's 100KB image model trains in 4 minutes

Training an AI to draw your specific cat usually involves gigabytes of weights, expensive GPU hours, and a lot of frustration when the model simply overfits to the training photos.

Nvidia, alongside researchers from Tel-Aviv University, just introduced Perfusion. It is a new text-to-image personalization method that shrinks custom training down to an astonishing 100KB file size and requires only about 4 minutes to train.

Locking concepts without breaking the scene

When AI art tools learn new concepts, they typically suffer from catastrophic overfitting. The model memorizes the reference pictures so tightly that it can only spit out the same poses in the same lighting over and over.

Perfusion solves this with a mechanism called Key-Locking. Instead of brute-forcing the entire model, it ties a personalized object—like your pet cat Tom—to a broader category concept like "feline" during the generation process.

By anchoring Tom to general catness, the base model still knows how cats stretch, sleep, or sniff flowers across various scenes. The AI only has to hold onto Tom's specific identifying traits while leaving the rest of the canvas completely flexible.

It also allows users to merge multiple personalized concepts in a single prompt with natural interactions, like having that specific cat sit on a specific custom chair without visual artifacts.

100KB takes on the multi-gigabyte giants

To understand why 100KB matters, look at current tuning methods.

LoRA files inside the Stable Diffusion ecosystem frequently range from dozens of megabytes up to over 1GB. DreamBooth models, widely considered the gold standard for visual fidelity, routinely weigh over 2GB. Textual inversion embeddings are light, but they often struggle with complex visual accuracy.

Perfusion operates on a fraction of that footprint because it only updates the exact cross-attention components needed to guide the image, rather than overhauling the whole network.

Even better, Nvidia baked in an inference-time slider. Users can dynamically balance visual fidelity against prompt alignment without retraining. You can decide whether you want the output to match the original object more strictly, or give the prompt more creative freedom.

Nvidia claims Perfusion outperforms fine-tuning techniques used by Stable Diffusion v1.5, SDXL, and MidJourney in edition efficiency. The team has published the paper and promised code will drop soon.

Why it matters

Nvidia is dominating the hardware gold rush of generative AI, but Perfusion shows it is equally dangerous on algorithmic efficiency.

For creators and app builders, multi-gigabyte models are a logistical nightmare. Storing personalized checkpoints for millions of users turns cloud hosting bills into a money pit. Squeezing a concept into 100KB makes personalized image models portable enough to store on edge devices, send over messaging apps, or swap instantly in memory.

We think this is one of the smartest research pivots in generative media. Instead of fighting over base model sizes, Nvidia is fixing the hyper-specialization layer that actual businesses care about.

The catch? The code is not out yet. While the Pareto-front flexibility sounds incredible on paper, community benchmarks will determine whether Key-Locking holds up across obscure objects and stylized art. But if it delivers, the era of hoarding 2GB checkpoints might be coming to an end.