Meta unveils Emu Video and Emu Edit for precise image editing and video generation

Meta just dropped two research breakthroughs aimed at fixing generative AI's biggest headache: control.

Anyone who has tried editing an image with AI knows the drill. You ask to change a small detail, and the model redraws the entire scene, morphing faces and lighting. On November 16, Meta unveiled Emu Edit and Emu Video, two models built on top of its Emu foundational image model introduced in September.

Editing pixels, not the entire picture

Meta's premise for Emu Edit is straightforward: an editing tool shouldn't just create a plausible picture, it should alter only the pixels relevant to the request. If you ask to write "Aloha!" on a baseball cap, the cap itself and the background shouldn't change.

Emu Edit handles text-based requests for local and global edits, background removal, color swaps, and geometric adjustments. To pull this off, Meta trained it on a dataset of 10 million synthesized examples covering 16 computer vision tasks, each containing an input image, a task prompt, and the expected output. Meta says incorporating vision tasks allows the model to generalize well, even tackling tasks like inpainting with minimal examples.

A two-step recipe for video

Video generation usually requires messy, heavyweight systems. Meta's earlier Make-A-Video relied on a cascade of five separate models. Emu Video trims that bloat down to just two diffusion models.

It works in a "factorized" two-step process: first, it generates a still image from a text prompt; second, it uses both the prompt and that generated image to produce motion. The result is a 512x512 video running at 16 frames per second for four seconds, with tests reaching up to eight seconds. It can also animate existing still photos from user prompts.

According to Meta, human evaluators heavily favored Emu Video. Compared to Make-A-Video, 96% preferred its quality and 85% found it more faithful to prompts. Meta also noted that in blind tests, Emu Video beat tools like Pika Labs across quality metrics in over 95% of cases and outperformed Google's Imagen Video.

Why it matters

Neither model is an open, standalone product yet, but this is a massive signal for how consumer AI will look inside Instagram and WhatsApp. Instead of hunting through GIF libraries or wrestling with Photoshop layers, users will soon generate animated stickers on the fly or adjust family photos with conversational text prompts.

We think Meta made a really smart architectural bet here. Cutting the video pipeline from five models to two shows that factoring the problem into distinct stages produces cleaner motion with less compute overhead. Meanwhile, Emu Edit's obsession with preserving untouched pixels addresses the exact reason creators still avoid generative image editors. But synthetic training data only gets you so far; whether these models can handle messy, real-world camera artifacts without artifacting remains to be seen once they leave the lab.