strong-followup
Lumina-Image 2.0: A Unified and Efficient Image Generative Framework
Qi Qin, Le Zhuo, Yi Xin · Shanghai AI Laboratory · 2025-03 · arXiv:2503.21758
Why it matters
An efficiency-oriented unified text-to-image framework using a Gemma-2 text encoder, reported to beat SD3 and Janus-Pro on several axes at lower cost. A common open efficiency baseline.
What this paper does
read: full textBefore this
Lumina-Next and comparable DiT-based text-to-image systems injected text conditioning through zero-initialized gated cross-attention, treating text embeddings as a fixed external feature rather than letting text and image tokens interact directly. Captions used to train these systems came from generic vision-language models not tailored to text-to-image data, producing captions that were inaccurate or insufficiently detailed.
The problem
Cross-attention conditioning limits the efficiency of multimodal fusion and constrains task expansion, while the lack of a captioning system built specifically for text-to-image data caps how well a model can learn prompt alignment.
The idea
Treat text and image tokens as one joint sequence processed by self-attention instead of routing text through a separate cross-attention branch, and pair this with a captioner purpose-built for text-to-image training data rather than a generic VLM captioner.
How it works
Unified Next-DiT concatenates text tokens from a frozen Gemma-2 2B encoder with image tokens and processes them jointly through single-stream self-attention, so text and image tokens attend to each other directly instead of through gated cross-attention. Generation happens in VAE latent space, and the model is trained with a flow-matching objective that regresses the velocity field between noise and data rather than predicting noise or clean data directly. UniCap, a fine-tuned Qwen2-VL-7B, generates multi-granularity, multi-perspective, multilingual captions for the roughly 110M-image training set, and training proceeds in three progressive stages at 256px on about 100M images, 1024px on about 10M images, and a final high-quality tuning stage on about 1M images. The model totals 2.6B parameters, trained for roughly 591 A100 GPU-days. At inference, CFG-Renormalization rescales velocity magnitude to avoid over-saturation at high guidance scales and CFG-Truncation skips conditional evaluation after early timesteps for about a 20% speedup.
Evidence
Lumina-Image 2.0 reaches GenEval 0.73 and DPG-Bench 87.20 (best among the models it compares against), and 0.8211 on T2I-CompBench color and 0.6028 on shape. In human preference arenas as of February 2025 it scores 982 Elo on Artificial Analysis (behind FLUX Pro at 1122, ahead of DALL-E 3 at 970) and 969 overall / 1031 alignment on Rapidata (second to FLUX Pro's 1040), and its AGI-Eval score of 0.4545 is well above Lumina-Next's 0.3229. All of this comes from a 2.6B-parameter model, smaller than many of the systems it is compared against.
Limitations
The authors concede the model struggles to consistently generate correct results for complex and diverse structures such as human bodies and for rare concepts, shows disordered details in dense textures, and fails at accurately rendering long and complex text.
Why it matters
It establishes joint self-attention over concatenated text-image tokens, rather than cross-attention, as a cheaper and more effective way to fuse modalities in a DiT, and demonstrates that a captioning system tailored to text-to-image training data measurably improves prompt alignment at a fixed parameter budget.
Abstract, in the authors' own words
We introduce Lumina-Image 2.0, an advanced text-to-image generation framework that achieves significant progress compared to previous work, Lumina-Next. Lumina-Image 2.0 is built upon two key principles: (1) Unification - it adopts a unified architecture (Unified Next-DiT) that treats text and image tokens as a joint sequence, enabling natural cross-modal interactions and allowing seamless task expansion. Besides, since high-quality captioners can provide semantically well-aligned text-image training pairs, we introduce a unified captioning system, Unified Captioner (UniCap), specifically designed for T2I generation tasks. UniCap excels at generating comprehensive and accurate captions, accelerating convergence and enhancing prompt adherence. (2) Efficiency - to improve the efficiency of our proposed model, we develop multi-stage progressive training strategies and introduce inference acceleration techniques without compromising image quality. Extensive evaluations on academic benchmarks and public text-to-image arenas show that Lumina-Image 2.0 delivers strong performances even with only 2.6B parameters, highlighting its scalability and design efficiency. We have released our training details, code, and models at https://github.com/Alpha-VLLM/Lumina-Image-2.0.
Research line
Reported results
| Benchmark | Value | Guidance | Budget | Source |
|---|---|---|---|---|
| DPG-Bench | 87.2 | — | — | Table 4 |
| GenEval | 0.73 | — | — | Table 4 |
Design-axis choices
Architecture
Conditioning & control
Builds on
Nothing recorded yet.
Built on by
- SVG-T2I: Scaling Up Text-to-Image Latent Diffusion Model Without Variational Autoencoder uses_architecture_from this — SVG-T2I's paper states it uses "the Unified Next-DiT (Qin et al., 2025) architecture as our backbone," which the citation identifies as Lumina-Image 2.0's single-stream joint-sequence DiT variant, combining Gemma2-2B text tokens with image tokens as one shared-weight sequence rather than MM-DiT's two-weight-set design.
This is a strong-followup entry — a short-form summary. Full paper-page explanations (before/problem/ core idea/representation/architecture/objective/conditioning/training/inference/results/ ablations/limitations) are written for landmark and core papers first; see PROJECT_STATE.md for the schedule.