NVIDIA's LoRAWeave edits images using before-and-after pairs

Prompting an AI to edit a picture just the way you want often feels like a guessing game. You can type dozens of adjectives like "claymation texture" or "rough architectural sketch," but the model usually delivers something slightly off.

NVIDIA introduced LoRAWeave, an image editing approach that ditches prompt gymnastics entirely. Instead of telling the model what to do with words, you simply show it.

Visual analogies over text prompts

Most reference-based image tools only take a single target or style reference. LoRAWeave asks for three images in total.

The first two form an example pair: image A is the "before" shot, and image A' is the "after" edit. The third image is target B, which is what you actually want to modify.

The system's job is to solve a visual analogy: A is to A' as B is to B'. It looks at the transformation between the first two images, figures out the underlying logic, and replicates the exact same edit onto your new target image.

For instance, if you feed it a real photo and its architectural sketch counterpart, you can drop in a fresh street view and get the exact same sketch style. Show it a bird turning into Studio Ghibli art, and it can transform your cat into that exact anime look. Show it a portrait turned into claymation, and it turns another person into clay as well.

It takes an extra setup step, but it keeps the visual translation remarkably consistent.

Composing a space of LoRAs

Visual analogies sound intuitive, but making a neural network generalize across unseen transformations is notoriously tricky.

Previous efforts usually relied on a single Low-Rank Adaptation (LoRA) module to learn visual edits. The catch is that a single module simply doesn't have enough capacity to capture the infinite variety of real-world style shifts.

NVIDIA's researchers solved this by introducing a learnable basis of LoRAs—essentially a foundational palette of transformation primitives.

They paired this with a lightweight encoder. When you upload your before-and-after pair, the encoder analyzes the differences and dynamically assigns weights to those basis LoRAs in a single inference pass. Instead of forcing one model to know everything, it dynamically builds a tailored LoRA on the fly.

Why it matters

We think LoRAWeave hits on an important insight: natural language is fundamentally flawed when it comes to describing nuanced visual edits.

For designers, digital painters, and creators, nuances like a subtle color grading curve, paper grain, or brush texture are nearly impossible to articulate in a text prompt. Being able to define an edit simply by providing two screenshots removes the prompt-engineering barrier and turns visual workflow duplication into a single click.

It also nudges the image generation industry toward real multimodal context. Instead of treating text prompts as the master interface, models should naturally understand visual demonstrations as direct commands.

The main friction right now is the data requirement. If you don't already have an accurate before-and-after pair lying around, creating that initial example is extra work. But for recurring, highly specific styling pipelines, this three-image analogy approach looks far more reliable than rolling the dice on prompt words.