AI Paper Atlas

Research map / Vision-language and multimodal / Multimodal LLMs

UniF$^2$ace: A $\underline{Uni}$fied $\underline{F}$ine-grained $\underline{Face}$ Understanding and Generation Model

Junzhe Li, Sifan Zhou, Liya Guo, Xuerui Qiu, Linrui Xu, TingTing Long, Chun Fan, Ming Li, Hehe Fan, Jun Liu, Shuicheng YAN

ICLR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Interleaved image-text generation with MLLMs

Read the paperCheck your idea against 137,217 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes UniF^2ace, a unified model for fine-grained face understanding and generation, addressing fragmented face-domain methods that handle understanding and generation separately.

Problem

Existing face-domain research is fragmented: methods fail to unify face understanding and generation within a single model, hindering progress toward unified face intelligence.

Approach

A unified multimodal model specialized for fine-grained face understanding and generation in one architecture, targeting the fragmentation of prior face-specific methods. Details beyond the abstract are not available.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add score-matching loss to masked generative loss (D3Diff, alpha 0.01) (ablation)UniF2ace · UniF2aceD-1M test (generation)FID: 77.463 -> 66.005Improved
use score loss only instead of masked loss only (ablation)UniF2ace · UniF2aceD-1M test (generation)FID: 77.463 -> 69.694Improved
vary D3Diff weight alpha from 0.01 to 0.001 (ablation)UniF2ace · UniF2aceD-1M test (generation)FID: 66.005 -> 72.736Hurt
add CLIP and Face experts in sequence-level MoE for understanding (ablation)UniF2ace · UniF2aceD-1M test (understanding)Desc-GPT: 5.21 -> 6.02Improved
add token-level MoE (ablation)UniF2ace · UniF2aceD-1M test (generation)FID: 72.877 -> 67.415Improved
add sequence-level MoE (ablation)UniF2ace · UniF2aceD-1M test (generation)FID: 72.877 -> 69.312Improved
combine token-level and sequence-level MoE (ablation)UniF2ace · UniF2aceD-1M test (generation)FID: 72.877 -> 66.005Improved
raise sequence-MoE Top-K from 1 to 2 (ablation)UniF2ace · UniF2aceD-1M test (generation)FID: 74.914 -> 66.005Improved
raise sequence-MoE Top-K from 2 to 4 (ablation)UniF2ace · UniF2aceD-1M test (generation)FID: 66.005 -> 63.632Mixed
add D3Diff loss to 1.3B baseline (cost) (ablation)baseline · UniF2aceD-1M test (generation)VLM-score: 79.698 -> 84.432Improved
add MoE architecture (1.3B to 1.8B params) (ablation)baseline · UniF2aceD-1M test (understanding)Desc: 4.958 -> 5.885Improved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy