ByteDance teaches image models left from right with spatial rewards

ByteDance researchers introduced a spatial reward modeling framework designed to fix one of image generation's most annoying flaws: mixing up left, right, front, and back. If you have ever burned through dozens of generations just trying to get an AI model to place an object behind another, this project is built for you.

What it actually fixes

Modern text-to-image models deliver stunning fidelity, but their spatial reasoning falls apart as prompts grow complex.

Take a crowded dining table prompt: three candles in a row with the tallest on the left and shortest on the right, a fruit bowl behind the middle candle, plates in front of the side candles, and a wall painting centered above the bowl. A standard model like Flux usually scrambles these relationships.

With ByteDance's framework, the generator actually places each element where it belongs. The team showed similar success across other messy setups, such as positioning armchairs, side tables, floor lamps, and teacups, or sorting out cars, traffic cones, backpacks, and water bottles.

How it works

Instead of retraining a massive base generator from scratch, the researchers used reward modeling and reinforcement learning.

They first created SpatialReward-Dataset, a collection of over 80,000 preference pairs. For each prompt, they paired a correct layout with an intentionally tampered, spatially incorrect image.

Using these pairs, they trained a specialized reward model called SpatialScore. The paper reports that SpatialScore evaluates spatial relationships more accurately than leading commercial models.

By plugging SpatialScore into an online reinforcement learning loop, the system guides generators to avoid positioning mistakes during sampling. The team has published the research paper and announced plans to open-source both the dataset and model weights.

Why it matters

Image generation has nailed aesthetics, but industrial design and marketing require precision. An image with the wrong layout is useless, no matter how photorealistic it looks.

We think targeted reward modeling is the right architectural bet here. Instead of burning compute to retrain frontier models, a specialized spatial critic lets existing open-weight models like Flux follow complex blueprints. Once ByteDance drops the weights, expect local creators and fine-tuners to jump on this immediately to cut down reroll fatigue.