i1: A Simple and Fully Open Recipe for Strong Text-to-Image Models

Modern text-to-image generation has entered an era of immense capability. Yet it suffers from a profound transparency crisis. While models like DALL-E 3 or FLUX.1 produce breathtaking imagery, they are often "black boxes." These are proprietary engines where training data and architectural nuances are guarded secrets. This opacity prevents scientists from understanding which design choices drive performance.

Current progress is bifurcated. You can choose state-of-the-art performance with zero reproducibility. Or you can choose "fully open" models. These rely on public weights, code, and data. However, they historically fall significantly short of closed-source counterparts. There has been a missing middle. We need high-performance models built using only publicly available ingredients to serve as a verifiable foundation for research.

The opacity of the state-of-the-art

Advancing text-to-image diffusion models—systems that generate images by iteratively reversing a noise process—is difficult. First, leading models rarely disclose their training data or full recipes. This makes it impossible to perform controlled ablations. An ablation is the process of systematically removing or changing parts of a system to see what they actually do. Without knowing if a model improved due to architecture or data, researchers are flying blind.

Second, even when models release open weights, they often bundle hundreds of variables into one "recipe." This bundling makes it nearly impossible to disentangle specific choices. Researchers cannot easily tell if performance comes from using multiple text encoders or how to balance different image data. Existing fully open models have struggled to bridge this gap. This leaves a performance chasm between what is possible and what is verifiable.

A recipe built on controlled ablation

Rather than proposing a radical new mathematical primitive, the authors of the i1 paper conducted a systematic investigation. They ran over 300 controlled experiments using 700,000 TPU v6e hours. Their goal was to decompose the text-to-image problem into manageable design choices. The resulting i1 model is a 3B-parameter model. It succeeds through the strategic selection of existing components rather than pure complexity.

The architecture, as illustrated in, relies on several key mechanical pillars:

1. Dual-Stream MMDiT Backbone: Instead of a single stream where text and image tokens are processed together, i1 uses a Multimodal Diffusion Transformer (MMDiT). This employs modality-specific parameters for text and image tokens. This allows for more specialized processing of each type of data.
2. Long Skip Connections: The authors incorporate "shortcuts" between early and late layers of the transformer. This design choice helps maintain structural integrity across the network.
3. Large Text Encoder Adapter: One significant finding involves how the model "listens" to text. Instead of using multiple expensive text encoders, the authors found a better way. They used a single strong encoder (T5Gemma-2B) paired with a larger, more expressive transformer adapter. This adapter replaces a simple MLP (a basic neural network layer).
4. Inference-Time Prompt Rewriting: To resolve a conflict between training and testing, the authors use a Large Language Model (LLM) to expand short user prompts. These are turned into descriptive paragraphs. This ensures the model encounters the long, detailed captions it saw during training.

Efficiency over complexity

The results of this systematic search suggest that scaling might be more about data orchestration than parameter count. The authors report that i1 achieves state-of-the-art performance among fully open models. It outperforms the best existing fully open model by 29.5 absolute percentage points on average across five major benchmarks .

Crucially, i1's 3B-parameter footprint allows it to compete with much larger models. For instance, the paper shows i1 performs competitively with the 12B-parameter FLUX.1 [Dev] and the 17B-parameter HiDream-I1 . This indicates that a well-tuned 3B model can match the utility of much heavier architectures. This is possible if data mixing and conditioning are handled correctly.

The authors also highlight the importance of data balancing. Through their experiments, they found that "equal weighting" is a powerful default. This means treating every curated dataset as equally important regardless of its raw size. This prevents a single massive dataset from drowning out specialized signals. Small, high-quality datasets for text-rendering or synthetic collections remain influential in the training mix.

Limits of the current regime

While i1 represents a leap for open research, it is not a perfect generator. The authors are transparent about several persistent failure modes. First, the model struggles with high-fidelity human synthesis in complex scenes. In group settings, it can produce distorted faces or malformed limbs [Figure 34a]. Second, it occasionally fails to grasp fundamental physical laws. One example involves failing to capture the correct behavior of a mirror reflection [Figure 34b].

Furthermore, the study's scope is bounded by its scale. Most controlled experiments were conducted on models of roughly 3B parameters or smaller. While the findings are compelling, the researchers note that scaling effects are unknown. It is unclear if these design preferences hold true for much larger models. Finally, the reliance on automated benchmarks means the model is excellent at "prompt following." However, its alignment with subjective human aesthetic preference has not been exhaustively measured.

The verdict: a foundation for open science

The i1 project argues that high-performance AI does not require proprietary secrecy. By providing the model weights, code, and the complete data processing pipeline, the authors move toward reproducible science.

The core takeaway for practitioners is that complexity is often a mask for suboptimal design. You do not need a dozen text encoders if you have a sufficiently expressive adapter. You do not need massive high-resolution datasets if you master low-resolution pre-training and prompt expansion. For the research community, i1 serves as a high-quality baseline. It finally allows researchers to ask specific questions about individual design components.

The code and model artifacts are reportedly available at the canonical links provided in the paper.

Figures from the paper