Lumos-X: Relate Any Identities with Their Attributes for Personalized Video Generation

TL;DR

Lumos-X targets the face-attribute alignment problem in multi-subject personalized video generation: when several characters appear together, current personalization models routinely swap a subject’s clothing onto another or attach the wrong attribute to the wrong identity. The paper attacks this on two fronts. On the data side, a curation pipeline assembles training tuples from independent source videos by using MLLMs to extract entity-attribute dependencies, then segments and re-composes subjects + attributes + clean backgrounds. On the modeling side, the Wan2.1-T2V DiT backbone is augmented with two new attention variants — Relational Self-Attention and Relational Cross-Attention — that combine position-aware embeddings with explicit subject-attribute coupling, enforcing intra-group cohesion (the same person stays bound to their own attributes) and inter-group separation (distinct subject clusters don’t bleed).

Key claims

The dominant failure mode for current personalized multi-subject video models is poor face-attribute alignment — existing methods have no explicit mechanism to ensure intra-group consistency (subject A’s attributes stay on subject A in every frame) [Abstract].
Addressing this requires both a face-attribute-aware data resource and explicit modeling — neither alone is sufficient [Abstract].
The data pipeline runs three steps over raw videos: (i) generate a caption + detect human subjects from extracted frames, (ii) retrieve entity words from the caption and match each subject to its attributes via MLLM inference, (iii) use the entity tags to localize and segment target subjects + objects and produce a clean background image [Fig. — Dataset construction pipeline].
The same pipeline doubles as a benchmark construction recipe — the extracted relational priors structure both training and evaluation [Abstract].
Lumos-X is built on Wan2.1-T2V; all condition images are encoded into image tokens via a VAE encoder, concatenated with the denoising video tokens, and fed into the DiT blocks [Fig. — Overview].
Within each DiT block, Relational Self-Attention and Relational Cross-Attention enable causal conditional modeling, enhance visual token representations, and “ensure precise face–attribute alignment” [Fig. — Overview].
The two new attention modules combine position-aware embeddings with refined attention dynamics to inscribe explicit subject-attribute dependencies, enforcing intra-group cohesion and amplifying separation between distinct subject clusters [Abstract].
Comprehensive benchmark evaluations report state-of-the-art performance in “fine-grained, identity-consistent, and semantically aligned” personalized multi-subject video generation, on the new benchmark built from the data pipeline [Abstract].

Method

The conditioning pipeline starts at the data layer. Raw videos go through a three-stage curation: (1) a caption is generated and human subjects are detected in extracted frames; (2) entity words are retrieved from the caption and an MLLM matches each subject in the visual stream to its corresponding attribute words; (3) the entity tags are used to localize and segment subjects + objects, and the residual is used to produce a clean background image. The outputs are (subject crop, attribute crop, clean background) triples that carry an explicit relational structure between identity and attribute.

The model is initialized from Wan2.1-T2V. At inference, condition images (subjects, attributes, optionally background) are encoded via the Wan VAE into image tokens, concatenated along the sequence dimension to the noisy denoising video tokens, and fed into the DiT stack. The architectural novelty is inside each DiT block: the standard self- and cross-attention are replaced with Relational Self-Attention and Relational Cross-Attention. These mix position-aware embeddings (encoding which subject cluster a given token belongs to) with attention masks/biases that strengthen attention within a subject-attribute group and weaken attention across groups. The net effect is that during denoising, the visual tokens for subject A’s region attend predominantly to subject A’s identity and attribute tokens, and away from subject B’s, even when both share the same frame.

Results

The project page reports SOTA on the authors’ new face-attribute-alignment benchmark for fine-grained, identity-consistent, semantically aligned multi-subject personalized video generation [Abstract]. Quantitative tables and baseline comparisons are reserved for the full arxiv PDF (2603.20192), which was not retrievable at filing time; this page captures what the project page advertises.

Why it’s interesting

Lumos-X is the fourth filed paper in 2026 that puts a custom identity-conditioning recipe on top of a Wan-family backbone, alongside ConsID-Gen: View-Consistent and Identity-Preserving Image-to-Video Generation (image↔geometry dual-stream over Wan2.1-1.3B), DreamID-Omni: Unified Framework for Controllable Human-Centric Audio-Video Generation (Syn-RoPE + Structured Captions over Ovi for joint A+V), and Wan-Animate: Unified Character Animation and Replacement with Holistic Replication (skeleton + face features over Wan2.2). Each picks a different conditioning primitive: ConsID-Gen adds geometry tokens from VGGT; DreamID-Omni binds identity-timbre with per-identity RoPE slots; Wan-Animate uses an explicit region-flag input; Lumos-X manipulates attention to enforce subject-attribute grouping. Together they suggest a small zoo of “frozen Wan + lightweight identity adapter” recipes is converging — a cluster worth a concept page once the recipes are easier to compare on a shared benchmark. The data-pipeline angle (“use MLLMs to infer relational priors over independent videos, then re-compose”) is also a concrete instance of the Synthetic Training Data pattern, applied to multi-subject personalization rather than to motion or action labels.