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

TL;DR

i1 is a 3B-parameter text-to-image diffusion model from Princeton (Zhuang Liu’s lab) trained as a controlled study of the T2I design space — 300+ ablations, 700K+ TPU v6e hours — that culminates in a single, fully-open recipe (weights + code + data pipeline) competitive with leading open-weight models. The model uses a dual-stream MMDiT backbone with long skip connections, a single T5Gemma-2B text encoder paired with a 2-block transformer adapter (no AdaLN noise conditioning), equally-weighted mixing of 12 public datasets with long Qwen3-VL-30B-A3B captions, and inference-time prompt rewriting. Headline result: on GenEval / DPG-Bench / PRISM / CVTG-2K / LongText-Bench, i1-3B beats the previous best fully-open model by 29.5 absolute percentage points on average and is competitive with closed and open-weight-only systems like Z-Image (6B), Qwen-Image (20B), and HiDream-I1-Full (17B). The paper’s value is less the model and more the cleanly-attributed inventory of which design choices actually matter — useful as a reference baseline for follow-on work.

Key claims

A dual-stream MMDiT backbone has the best performance-parameter trade-off among cross-attention, single-stream, and dual-stream families across model sizes [Backbone Architecture section, design table].
Long skip connections improve the dual-stream backbone’s performance-parameter trade-off and are kept in the final i1 [Backbone Architecture section].
Encoder-decoder T5Gemma-2B is the strongest text encoder in the comparison, beating CLIP-style and decoder-only LLM/VLM text encoders for T2I training [Text and Noise Conditioning section, text-encoder table].
The gain from using multiple text encoders can be largely captured by increasing a single encoder’s adapter capacity (concat two copies of T5Gemma-2B features via two MLP adapters ≈ adding a second encoder; a shared MLP does not) — adapter capacity, not encoder count, is the active variable [Text and Noise Conditioning section, two-MLP repeat ablation].
AdaLN conditioning on pooled text embeddings or timestep embeddings adds substantial parameters but only marginal benefit; removing AdaLN in single-stream improves DPG by +1.49 / PRISM by +3.4 / LongText by +0.097 at ~0.25B fewer params [Impact of text and noise conditioning table].
Training-caption length and test-prompt length must be aligned: training only on long captions and using LLM-based prompt rewriting on short prompts at inference is the strongest configuration (GenEval 0.73 vs 0.47 with short-only training, 0.17 with short-test on long-train) [Synthetic Captions table].
The choice of synthetic-captioner VLM substantially affects downstream T2I performance — Qwen3-VL-30B-A3B captions give the strongest result among the compared captioners [Synthetic Captions section].
Equal weighting across 12 curated public datasets is a strong default for dataset mixing; capping at lower thresholds (more even weights) consistently beats higher thresholds [Threshold-based weighting table].
Among real-image datasets, ImageNet-22K and YFCC are the strongest single-dataset sources; iNaturalist is the weakest and dropping it improves all three benchmarks [Single-dataset and Removing weakest datasets tables].
Real, synthetic, and text-rendering image subsets are all individually necessary — removing any one degrades at least one benchmark [Dataset Mixing section].
With diverse dataset mixing, repetition is cheap: subsampling each dataset to 0.4M images (4.4M unique seen vs 88.1M) gives only marginal performance loss [Subsampling table].
Text-rendering capability improves substantially with 512-resolution training even when text-rich datasets are not in the mix at that stage, suggesting high-resolution training does not need the full breadth of the low-res pre-training mixture [512-resolution benchmark performance section].
On the five benchmarks, i1-3B (3B) reports GenEval 0.84 / DPG 86.73 / PRISM 70.1 / CVTG-2K 0.8531 / LongText 0.922 — beating the next-best fully-open model (BLIP3o-N-G-G at GenEval 0.90 but PRISM 57.5 / LongText 0.114) by 29.5 absolute percentage-point average across the five benchmarks, and matching weight-only models like Z-Image 6B and Qwen-Image 20B on PRISM/CVTG/LongText [final benchmark table].

Method

i1 follows a single-variable-at-a-time controlled-experiment protocol: every ablation starts from the same XL/2-sized 256-resolution baseline (LightningDiT-style cross-attention backbone, QK-norm, long skip connections, T5Gemma-2B text encoder, FLUX.2 VAE, 500K iters × bs 512 × lr 1e-4, 250-step Euler + CFG=12). Modifications are not accumulated across experiments — each ablation isolates one design choice. The final recipe combines the winning choices: a dual-stream MMDiT backbone with long skip connections; a single T5Gemma-2B text encoder feeding a 2-block transformer adapter; no AdaLN; both sinusoidal and RoPE positional embeddings; shared sandwich normalizations across text and image streams. The data side combines 12 curated public datasets (ImageNet-22K, YFCC, etc.) equally weighted, with long synthetic captions from Qwen3-VL-30B-A3B and inference-time prompt rewriting via a single meta-prompt.

Training is staged across resolutions: 256-resolution pretrain (162.9M images, 2M steps, bs 512, 383 TPU v5p-128 hours) → 512-resolution (9.7M filtered images, 500K steps, 174 hours) → 1024-resolution (4.3M filtered images, 300K steps, bs 128, timestep shift 3.33, 151 hours). At inference, CFG=12 with Rescale CFG strength 1.

Results

On GenEval / DPG-Bench / PRISM / CVTG-2K / LongText-Bench, i1-3B reports 0.84 / 86.73 / 70.1 / 0.8531 / 0.922. Against fully-open models (weights + data + code), the previous-best entries are BLIP3o-N-G-G (GenEval 0.90 but CVTG 0.2442, LongText 0.114) and DeCo (GenEval 0.86, LongText 0.003) — i1 wins on every benchmark except GenEval where BLIP3o-N variants edge ahead, and beats the best fully-open model by 29.5 absolute percentage points averaged across the five. Against open-weight-only models, i1 (3B) is competitive with FLUX.1 [Dev] (12B, GenEval 0.66 / LongText 0.607), HiDream-I1-Full (17B, PRISM 66.1 / CVTG 0.7738), and Z-Image (6B, PRISM 74.2 / CVTG 0.8671). Against closed API models, i1 is close to GPT Image 1 [High] on LongText (0.922 vs 0.956) and CVTG (0.8531 vs 0.8569) while trailing on GenEval (0.84 vs 0.84). The text-encoder ablation shows the encoder-decoder T5Gemma-2B beating decoder-only LLMs/VLMs across DPG / PRISM / LongText. The text-encoder-adapter ablation shows a 2-block transformer adapter raising DPG 84.66 → 86.33 and LongText 0.211 → 0.414 over the MLP-adapter baseline, with multi-encoder gains shrinking once adapter capacity is increased. The dataset-mixing ablation shows equal weighting beating threshold-based weights at all four tested thresholds.

Why it’s interesting

i1 is the strongest current data point for the “open-recipe-as-artifact” mode of Open foundation-model releases — the value is less the 3B model than the carefully-attributed inventory of which choices in a modern T2I stack actually move the needle, with weights, JAX training code, PyTorch inference, dataset and data-processing pipeline all released. It joins the dual-stream MMDiT cluster (Dual-stream diffusion transformer) with the strongest controlled head-to-head among cross-attention / single-stream / dual-stream backbones filed to date — and adds a negative result that contrasts with Z-Image’s (Z-Image: An Efficient Image Generation Foundation Model with Single-Stream Diffusion Transformer) and Ideogram 4.0’s (Ideogram 4.0 Technical Details: Open model at the forefront of design) single-stream wins, suggesting the dual-vs-single question is sensitive to scale, text-encoder family, and adapter design. Several findings line up with or sharpen earlier results: the text-encoder verdict overlaps with the T5Gemma 2 push (T5Gemma 2: Seeing, Reading, and Understanding Longer) and the encoder-decoder Gemma adaptation (Encoder-Decoder Gemma: Improving the Quality-Efficiency Trade-Off via Adaptation) by independently picking T5Gemma-2B over decoder-only LLMs at this scale; the “long-caption train + short-prompt rewrite at inference” recipe complements the structured-caption story in Generating an Image From 1,000 Words: Enhancing Text-to-Image With Structured Captions; the no-AdaLN finding aligns with Z-Image’s lightweight single-stream success and with the broader Diffusion training efficiency thread of removing components that buy parameters but not quality.