The "two skins" problem of unified models has been debated for two years: visual understanding and image generation each use their own representation and training objective, so sharing a backbone is unification in name only. Shanghai AI Laboratory has put Intern Lumina U2 on Hugging Face with a fully discrete answer: image, video, and 3D inputs are all encoded into multi-codebook discrete tokens and trained with text in one optimization space under a single masked-diffusion objective (paper: https://arxiv.org/abs/2610.20818).

One Visual Position, Eight Codebooks

The old weakness of single-codebook discrete tokens is limited representational capacity—each position compresses into one code, starving detail and semantics. U2 assigns each visual position 8 complementary codebooks (4096-entry vocab each, FSQ quantization). The visual tokenizer is Apple's open-source AToken, encoding images, videos, and 3D into a shared 4D latent space, kept frozen throughout training. The backbone handles this with restraint: at input, the 8 code embeddings are concatenated and projected into a single token, so sequence length does not grow with codebook count; at output, a lightweight multi-codebook AR head predicts codes sequentially within each position, while the diffusion backbone refines positions in parallel. Positional encoding stays with the backbone's native 1D RoPE, with spatial structure marked by structural delimiters.

A 1.4B-Activated MoE Base

The backbone converts from LLaDA2.0-mini: 16.9B total parameters, only 1.4B activated per token. Of its 20 Transformer layers, the first is dense; the remaining 19 are MoE layers with 256 routed experts plus one shared expert each, activating 8 routed experts per token. Only 0.65B parameters of visual interfaces were added, inheriting language capability without a second autoregressive-to-diffusion conversion.

Benchmarks: 6/10 First in Understanding, Closing on Qwen-Image

In official evaluations, U2 with 1.4B activated ranks first on six of ten image-understanding benchmarks, with the clearest leads on CharXiv (85.9), HallusionBench (64.6), MathVision (36.5), and DynaMath (63.6); it beats same-backbone LLaDA2.0-Uni on eight of ten. It tops all three video benchmarks (NVIDIA run: Video-MME 56.7, MLVU 64.9; Ascend run: MVBench 61.8) and scores 50.5 total on 3D MM-Vet, beating the 3D specialists PointLLM and ShapeLLM. On generation, among unified models it posts the best DPG-Bench (87.99), UniGenBench (86.11), and TIIF-Bench (87.1), matching the best GenEval at 89—remarkable given Qwen-Image activates roughly 14x more parameters per token. The shortcoming is plain: OneIG-Bench text rendering at 20.11 trails LLaDA2.0-Uni's 31.23, and sample diversity runs low.

4.45x Faster Image Generation via Confidence Plateaus

Diffusion decoding is expensive because of iterative refinement. The team found visual confidence is naturally low—its median crawls only from 0.046 to 0.127 over 96 steps—so Fast-dLLM-style absolute-threshold commit rules do not apply. They switched to temporal signals: positions entering confidence plateaus commit early, Halton low-discrepancy ordering spreads commits evenly, and V-verify drift tests decide when to refresh the cache. Stacked with CFG batching, AR-head KV caching, and hidden-state caching, production-path image latency drops from 159.9s to 35.9s—a 4.45x end-to-end speedup—while GenEval actually rises from 86.6 to 89.3.

Full-Stack on 1,024 Ascend 910B, Apache 2.0

The engineering side deserves note: training, inference, and evaluation were fully adapted on 1,024 Ascend 910B NPUs, with a smaller-budget NVIDIA reference run for cross-platform validation—Ascend is the primary training platform, not a backup. Code is open-sourced on GitHub under Apache 2.0, weights on Hugging Face. "First fully discrete multi-codebook diffusion language model" remains the team's own claim, and the text-rendering gap is real—but this release, bundling representational capacity, task unification, and a domestic compute stack in one open package, gives the unified-model route a reference implementation worth reproducing.