Intern Lumina U2 star logo Intern Lumina U2
Report Code Model

Intern Lumina U2

A Multi-Codebook Diffusion Large Language Model for Omni-Visual Understanding and Image Generation

Intern Lumina U2 unifies language, images, video, and 3D for text question answering, image understanding and editing, image generation, video understanding, and 3D understanding.

Explore demos ↓ View results ↓
Intern Lumina U2 capabilities and benchmark overview
Figure 1. Capabilities across image generation and editing, image understanding, video understanding, and 3D understanding.

Experimental Results

Task familyBenchmarkInternVL-ULLaDA2.0-UniLLaDA-oIntern Lumina U2
Documents and charts
Document & chartChartQA ↑76.680.187.986.2
Document & chartCharXiv-DQ ↑53.368.469.885.9
Perception and hallucination
PerceptionHallusionBench ↑44.850.247.464.6
PerceptionMME-C ↑279587527667
PerceptionRealWorldQA ↑56.466.760.862.9
Knowledge and mathematical reasoning
KnowledgeMMMU ↑54.750.144.952.2
KnowledgeMMMU-Pro ↑20.834.028.338.5
Math reasoningMathVista ↑55.868.166.168.0
Math reasoningMathVision ↑22.126.715.736.5
Math reasoningDynaMath ↑46.9746.8940.5863.6
Generation and editing
Image generationGenEval (0–100) ↑85898689
Image generationDPG-Bench ↑85.1887.7687.0487.79
Image generationTIIF (long) ↑73.9——87.1
Image generationWISE ↑0.460.68—0.56
Image editingImgEdit (1–5) ↑3.673.92—3.82
Video understanding
Compared againstLLaVA-MiniVideoLLaMA2-7BLLaDA-VIntern Lumina U2
VideoVideo-MME ↑—46.656.152.6
VideoMLVU ↑42.8—59.562.1
VideoMVBench ↑44.554.6—61.8
3D understanding
Compared againstPointLLM-7BShapeLLM-7B—Intern Lumina U2
3D3D MM-Vet (total) ↑41.247.4—50.5

Scores are the released Ascend-trained checkpoint; baseline numbers are taken from the tech report. GenEval is on a 0–100 scale. Video and 3D are compared against specialist models, named in the table where they apply. Video-MME is without subtitles and MLVU is m-avg. A dash means the model does not report that benchmark; Video-LLaVA-7B (Video-MME 39.9, MVBench 43.1) is in the report but does not fit the three baseline columns here.

Full comparison tables are in the technical report.

Method.

A unified architecture
for omni-visual tasks.

Intern Lumina U2 unified multi-modal architecture
Figure 2. Architecture overview. The encoder and two task heads share a MoE Diffusion LLM backbone across visual understanding, image generation, and editing.

Discrete VQ levels for visual tokens.

Comparison of universal, separate, and multi-codebook visual representations
Figure 3. Representation design. The proposed design uses discrete VQ levels to represent each visual position.

Parallel positions, sequential codebooks.

Multi-codebook autoregressive head for training and inference
Figure 4. Prediction head. Spatial positions are processed in parallel and codebooks are decoded in sequence.

Resources

Technical report Code Model 
@article{internluminau2,
  title   = {Intern Lumina U2: A Multi-Codebook Diffusion Large Language Model for
             Omni-Visual Understanding and Image Generation},
  author  = {{Intern Lumina U2 Team, Shanghai AI Laboratory}},
  journal = {arXiv preprint arXiv:2610.20818},
  year    = {2026}
}