Twins: Learn to Predict Unified Representations with Focal Loss

Kaixiong Gong, Xin Cai, Bin Lin, Hao Wang, Yunlong Lin, Mingzhe Zheng, Bohao Li, Jian-Wei Zhang, Miles Yang, Zhao Zhong, Liefeng Bo, Xiangyu Yue
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:35888-35902, 2026.

Abstract

Unified multimodal models seek a shared visual token space that supports both multimodal understanding and image generation. Discrete methods unify the interface via a shared codebook, whereas continuous pipelines often rely on two disparate representations—semantic features (e.g., ViT) for understanding and low-level latents (e.g., VAE) for synthesis—resulting in mismatched latent spaces. We propose Twins, a unified continuous token space formed by channel-wise concatenating ViT and VAE features on the same token grid, so the sequence length is unchanged and attention cost does not increase. However, jointly modeling Twins in a Diffusion Transformer exposes a severe optimization imbalance: the model fits the ViT component well but struggles to match the VAE latent distribution. We trace this imbalance to three sources of heterogeneity: frequency bias, intrinsic dimensionality, and condition-aligned vs condition-independent uncertainty. To address it, we adapt a focal regression objective for flow matching that upweights large-error VAE dimensions, better balancing optimization across the ViT and VAE components. On ImageNet, this yields up to $10.57$ gFID gain over naive MSE loss without classifier-free guidance. Twins also performs competitively on multimodal understanding benchmarks and improves reconstruction fidelity, narrowing the gap between understanding- and generation-oriented representations.

Cite this Paper

BibTeX
@InProceedings{pmlr-v306-gong26b, title = {Twins: Learn to Predict Unified Representations with Focal Loss}, author = {Gong, Kaixiong and Cai, Xin and Lin, Bin and Wang, Hao and Lin, Yunlong and Zheng, Mingzhe and Li, Bohao and Zhang, Jian-Wei and Yang, Miles and Zhong, Zhao and Bo, Liefeng and Yue, Xiangyu}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {35888--35902}, year = {2026}, editor = {Zhang, Tong and Dudik, Miroslav and Jaggi, Martin and Agarwal, Alekh and Li, Sharon and Schuurmans, Dale and Zhu, Jerry and Berkenkamp, Felix and Dong, Hanze and Bietti, Alberto}, volume = {306}, series = {Proceedings of Machine Learning Research}, month = {06--11 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v306/main/assets/gong26b/gong26b.pdf}, url = {https://proceedings.mlr.press/v306/gong26b.html}, abstract = {Unified multimodal models seek a shared visual token space that supports both multimodal understanding and image generation. Discrete methods unify the interface via a shared codebook, whereas continuous pipelines often rely on two disparate representations—semantic features (e.g., ViT) for understanding and low-level latents (e.g., VAE) for synthesis—resulting in mismatched latent spaces. We propose Twins, a unified continuous token space formed by channel-wise concatenating ViT and VAE features on the same token grid, so the sequence length is unchanged and attention cost does not increase. However, jointly modeling Twins in a Diffusion Transformer exposes a severe optimization imbalance: the model fits the ViT component well but struggles to match the VAE latent distribution. We trace this imbalance to three sources of heterogeneity: frequency bias, intrinsic dimensionality, and condition-aligned vs condition-independent uncertainty. To address it, we adapt a focal regression objective for flow matching that upweights large-error VAE dimensions, better balancing optimization across the ViT and VAE components. On ImageNet, this yields up to $10.57$ gFID gain over naive MSE loss without classifier-free guidance. Twins also performs competitively on multimodal understanding benchmarks and improves reconstruction fidelity, narrowing the gap between understanding- and generation-oriented representations.} }
Endnote
%0 Conference Paper %T Twins: Learn to Predict Unified Representations with Focal Loss %A Kaixiong Gong %A Xin Cai %A Bin Lin %A Hao Wang %A Yunlong Lin %A Mingzhe Zheng %A Bohao Li %A Jian-Wei Zhang %A Miles Yang %A Zhao Zhong %A Liefeng Bo %A Xiangyu Yue %B Proceedings of the 43rd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2026 %E Tong Zhang %E Miroslav Dudik %E Martin Jaggi %E Alekh Agarwal %E Sharon Li %E Dale Schuurmans %E Jerry Zhu %E Felix Berkenkamp %E Hanze Dong %E Alberto Bietti %F pmlr-v306-gong26b %I PMLR %P 35888--35902 %U https://proceedings.mlr.press/v306/gong26b.html %V 306 %X Unified multimodal models seek a shared visual token space that supports both multimodal understanding and image generation. Discrete methods unify the interface via a shared codebook, whereas continuous pipelines often rely on two disparate representations—semantic features (e.g., ViT) for understanding and low-level latents (e.g., VAE) for synthesis—resulting in mismatched latent spaces. We propose Twins, a unified continuous token space formed by channel-wise concatenating ViT and VAE features on the same token grid, so the sequence length is unchanged and attention cost does not increase. However, jointly modeling Twins in a Diffusion Transformer exposes a severe optimization imbalance: the model fits the ViT component well but struggles to match the VAE latent distribution. We trace this imbalance to three sources of heterogeneity: frequency bias, intrinsic dimensionality, and condition-aligned vs condition-independent uncertainty. To address it, we adapt a focal regression objective for flow matching that upweights large-error VAE dimensions, better balancing optimization across the ViT and VAE components. On ImageNet, this yields up to $10.57$ gFID gain over naive MSE loss without classifier-free guidance. Twins also performs competitively on multimodal understanding benchmarks and improves reconstruction fidelity, narrowing the gap between understanding- and generation-oriented representations.
APA
Gong, K., Cai, X., Lin, B., Wang, H., Lin, Y., Zheng, M., Li, B., Zhang, J., Yang, M., Zhong, Z., Bo, L. & Yue, X.. (2026). Twins: Learn to Predict Unified Representations with Focal Loss. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:35888-35902 Available from https://proceedings.mlr.press/v306/gong26b.html.

Related Material