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

Preliminary, partial results. Full comparison tables will appear in the upcoming technical report.

Task family	Benchmark	InternVL-U	LLaDA2.0-Uni	LLaDA-o	Intern Lumina U2
Visual reasoning & perception
Document & chart	ChartQA ↑	76.60	80.10	87.90	86.52
Document & chart	CharXiv-DQ ↑	53.30	68.40	69.80	83.65
Perception	HallusionBench ↑	44.80	50.20	47.40	62.15
Knowledge	MMMU-Pro ↑	20.80	34.00	28.30	36.13
Math reasoning	MathVision ↑	22.10	26.70	15.70	33.22
Math reasoning	DynaMath ↑	46.97	46.89	40.58	56.37
Temporal understanding
Video	Video-MME ↑	53.00	43.76	44.90	51.26
Video	MVBench ↑	59.61	48.97	47.16	59.74
Video	MLVU ↑	57.27	45.81	48.62	57.03
Generation & editing
Image generation	GenEval ↑	0.85	0.89	0.86	0.81
Image generation	DPG ↑	85.18	87.76	87.04	87.10
Image generation	TIIF Short / Long ↑	74.90 / 73.90	—	—	84.60 / 85.95
Image editing	ImgEdit ↑	3.67	3.92	—	3.8307

Method.

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.

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.

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

Demo Examples.

Text-to-Image Generation

✦

Image editing

✦

Image understanding

✦

3D understanding

✦

Video understanding

Resources

Intern Lumina U2

Experimental Results

Method.

A unified architecture
for omni-visual tasks.

Discrete VQ levels for visual tokens.

Parallel positions, sequential codebooks.

Demo Examples.

Text-to-Image Generation

Image editing

Image understanding

3D understanding

Video understanding

Coming soon.

Intern Lumina U2

Experimental Results

Method.

A unified architecturefor omni-visual tasks.

Discrete VQ levels for visual tokens.

Parallel positions, sequential codebooks.

Demo Examples.

Text-to-Image Generation

Image editing

Image understanding

3D understanding

Video understanding

Coming soon.

A unified architecture
for omni-visual tasks.