A shared interface lets image generators exchange denoising states
In plain words
Image generators start with noise and repeatedly change it into an image. InstantFusion is a method that lets one generator pass its partly formed image state to another, even when they store that state differently.
It translates the state through a shared representation trained to preserve the state and match how the generators change it.
In tests combining Qwen-Image with SD3 or FLUX.1, the combinations scored higher on an image-generation quality test than those target models alone. This finding applies to the tested combinations.
The research
Can image generators with different internal representations share useful intermediate states, or does handing a state from one model to another break the image as denoising continues?
Earlier methods link particular model pairs or individual states, leaving open whether one shared space can also connect how different generators update those states over time.
InstantFusion translates intermediate states through a shared representation, letting different image generators hand off denoising without a separate mapping for each pair.
The researchers trained a separate Latent AutoEncoder for each generator to place matched intermediate states in a common representation, while preserving the information needed to reconstruct them and aligning their denoising changes.
The generator backbones stayed frozen. On GenEval-v2, a test that scores the quality of generated images, Qwen-Image combinations scored above SD3 and FLUX. 1 alone.
This supports useful transfer in those tested combinations; objective ablations also showed different failures when dynamics alignment or reconstruction was removed.
The evidence covers four representative generators and the combinations reported here; it does not establish gains for every model or prompt.
The current method assumes a shared autoencoder and compatible time parameterization, and hybrid inference requires multiple frozen generators to be available, increasing memory and storage.
Analysis
Deep
- Authors
- Xuanlang Dai, Jiazi Bu, Yujie Zhou, Yuhong Liu, Beichen Zhang
- License
- arXiv non-exclusive