RecAReconstruction Alignment

ICLR 2026Accepted paper

Reconstruction
Alignment Improves Unified Multimodal Models

Improving image generation
through visual understanding.

A self-supervised post-training method for unified multimodal models. RecA reconstructs images from the model’s own visual understanding features—transferring visual understanding to text-to-image generation and image editing.

Paper ↗Code ↗Models ↗Try the demo ↗
A Van Gogh style painting of a cyberpunk city at sunrise, generated with RecA.
A cyberpunk sunrise, in the style of Van Gogh.GENERATED WITH RecA
Ji Xie1Trevor Darrell1Luke Zettlemoyer2XuDong Wang1,3,†
1 UC Berkeley2 University of Washington3 Duke University

† Corresponding author

02 / Self-supervised post-training

The image itself
is the supervision.

Reconstruction Alignment (RecA) trains a pretrained unified multimodal model to reconstruct an image from its own visual understanding features, using the original image as the target.

This semantic reconstruction objective enables understanding-to-generation transfer: better text-to-image generation and image editing, with the model’s original inference interface.

RecA training pipeline: an image is encoded into visual understanding features, which condition the model to reconstruct the original image.
Self-supervised post-training for unified understanding and generation. The reconstruction objective needs no image-specific captions.
01

Understand the image

Use the pretrained model’s visual understanding features as the conditioning signal.

02

Reconstruct its semantics

A semantic information bottleneck encourages learning meaningful visual structure through reconstruction.

03

Transfer to generation

Use the same inference interface for text-to-image generation and image editing.

Training scope. The reconstruction objective is caption-free. Models with shared understanding and generation parameters also retain image-to-text training to preserve understanding; decoupled understanding components can remain frozen. Evaluated on Show-o, Harmon, OpenUni, and BAGEL. Architecture-specific training guide ↗

The supervision gap

Images say more than captions.

Captions often leave out fine-grained attributes, spatial relationships, and visual details. Visual understanding features offer a denser source of supervision.

Comparison of sparse text-caption supervision with dense visual understanding features.
Visual features retain information that a text description can omit.

The alignment gap

Understanding is not yet generation.

A unified model can recognize a visual concept yet struggle to generate it. RecA connects these capabilities through semantic reconstruction.

Yellow broccoli example illustrating a gap between a model’s visual understanding and image generation.
The paper’s yellow-broccoli example illustrates this mismatch.

03 / Results

Better generation.
Stronger editing.

Improvements across distinct tasks and architectures, using visual understanding for generative model post-training.

Text-to-image generation

Harmon-1.5B → Harmon-1.5B + RecA

GenEval ↑0.73→0.86
DPG-Bench ↑80.93→87.21

A separate SFT → RecA recipe using GPT-4o-Image distillation data for supervised fine-tuning reaches 0.90 GenEval / 88.15 DPG-Bench.

Image editing

BAGEL → BAGEL + RecA

ImgEdit ↑3.38→3.75
GEdit-Bench-EN ↑6.94→7.27

This BAGEL experiment uses 10,000 unlabeled images, 1,000 training steps, and 27 A100 GPU-hours.

Full benchmark tables & per-category results
Full text-to-image benchmark comparison from the paper, including separate training recipes.
Generation benchmark comparison.
Per-category generation scores across Show-o, Harmon, OpenUni, and BAGEL.
Generation results across architectures and evaluation categories.
BAGEL image-editing results reported in the paper.
Image-editing evaluation.
Evaluation protocols & full experiments ↗

04 / Image editing

A new light. A different detail.

BAGEL-RecA editing examples. Drag the divider to compare the original image and the edited result.

Original image before: Shed neon light on the scene.
Edited result: Shed neon light on the scene.
OriginalRecA
01Shed neon light on the scene.
Original image before: Turn on the flashlight on the smartphone.
Edited result: Turn on the flashlight on the smartphone.
OriginalRecA
02Turn on the flashlight on the smartphone.
Original image before: Remove the feather on the hat.
Edited result: Remove the feather on the hat.
OriginalRecA
03Remove the feather on the hat.
Original image before: A man is wearing the clothes from the reference image.
Edited result: A man is wearing the clothes from the reference image.
OriginalRecA
04A man is wearing the clothes from the reference image.
Original image before: Add strawberries, blueberries, and banana slices on top of the pancake stack.
Edited result: Add strawberries, blueberries, and banana slices on top of the pancake stack.
OriginalRecA
05Add strawberries, blueberries, and banana slices on top of the pancake stack.
Original image before: Change the gingerbread cookies into soft plush toy figures, with fuzzy fabric texture.
Edited result: Change the gingerbread cookies into soft plush toy figures, with fuzzy fabric texture.
OriginalRecA
06Change the gingerbread cookies into soft plush toy figures, with fuzzy fabric texture.

Keyboard: focus a comparison and use the arrow keys. Try BAGEL-RecA ↗

05 / Citation

Build on this work.

If RecA is useful or inspiring for your research, please consider citing our ICLR 2026 paper. Abstract & publication details ↗

BibTeX ↓
@inproceedings{xie2026reconstruction,
  title={Reconstruction alignment improves unified multimodal models},
  author={Xie, Ji and Zettlemoyer, Luke and Wang, Xudong and others},
  booktitle={International Conference on Learning Representations},
  volume={2026},
  pages={120095--120137},
  year={2026}
}