PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Boya Zeng, Tianze Luo, Shu Pu, Jucheng Shen, Taiming Lu, Gabriel Sarch, Zhuang Liu

arXiv:2606.11289v1cs.CV

TL;DR

Opaque training data and recipes make it difficult to identify which modeling and data choices drive text-to-image progress, while fully open models lag behind leading systems. The paper uses controlled experiments to develop i1, a 3B-parameter model trained on public datasets, achieving state-of-the-art performance among fully open models across representative benchmarks.

  • Problem

    Undisclosed training data and recipes make it difficult to disentangle which modeling and data choices drive performance, while fully open models lag behind leading models.

  • Method

    The authors independently vary modeling, data, and inference choices in controlled experiments before combining improvements into i1.

  • Results

    i1 achieves state-of-the-art performance among fully open models at 1024 resolution and outperforms several larger open-weight-only models across representative benchmarks.

  • Takeaways & Limitations

    The released model, code, data recipe, and ablations provide a strong, reproducible baseline and reference point for open text-to-image research.

  • Takeaways & Limitations

    GenEval may not accurately reflect model capability because it can be misaligned with human judgment and inflated by BLIP3o-60K fine-tuning.

Abstract

from arXiv · show

Diffusion models have consistently driven progress in text-to-image generation. However, it is challenging to attribute recent progress to specific modeling and data choices: state-of-the-art open-weight models provide limited ablations, and do not disclose their training data and full training details. The research community needs fully open (weights, data, and code) models as a foundation for further research; yet existing fully open models still fall significantly short of leading models in performance. In this project, we conduct a systematic investigation of the modeling and data design choices in text-to-image diffusion training and inference with 300+ controlled experiments totaling 700K+ TPU v6e hours. Our experiments highlight several empirical findings (e.g., equal weighting is a strong default for mixing curated datasets) and simple design decisions (e.g., larger text encoder adapters improve performance with minimal added parameters) for training strong models. Guided by these insights, we train i1, a 3B-parameter text-to-image diffusion model using only publicly available datasets. i1 is competitive with leading models on five representative benchmarks (GenEval, DPG, PRISM, CVTG-2K, and LongText), and outperforms the best existing fully open model by 29.5 absolute percentage points on average. We provide the i1 checkpoints, training and inference code, and the data processing pipeline. Together, our findings and the i1 recipe establish a practical foundation for future open research in text-to-image diffusion models. Our code is available at https://github.com/zlab-princeton/i1.

1 Introduction

The introduction identifies limited openness and ablation evidence as barriers to understanding text-to-image progress, then presents controlled experiments that guide a simple, strong, fully open 3B-parameter model.

  • Motivation: Diffusion models have advanced text-to-image generation, but opaque training details and insufficient ablations hinder reproducibility and attribution of improvements.Leading models often withhold training data, full recipes, and thorough design analyses, while fully open models substantially lag in performance.
  • Experimental approach: Controlled experiments investigate text conditioning, noise and timestep conditioning, backbone architectures, and data design choices, primarily during 256×256 pre-training.The study begins from a simple baseline and systematically evaluates existing and new architectural and data designs.
  • Modeling findings: A single strong text encoder with a larger adapter, limited use of timestep or noise conditioning, and a dual-stream DiT with long skip connections form the key modeling choices.The larger adapter can outperform combining multiple text encoders, while timestep/noise conditioning and Adaptive Layer Normalization provide little benefit in this setting.
  • Data findings: Long captions strengthen models but hurt short prompts, synthetic captioner choice matters, and equal weighting across curated datasets is a strong default.Prompt rewrite at inference can mitigate the short-prompt weakness caused by long-caption training.
  • i1 model: 3B-parameter i1 is trained on publicly available datasets and achieves state-of-the-art performance among fully open models at 1024-resolution, surpassing larger open-weight-only models.The cited larger models include 17B HiDream-I1 and 12B FLUX.1 [Dev].
  • Contribution: i1 combines carefully selected modeling and data choices rather than major new network modules, while providing weights, code, datasets, and related research resources.Its design includes underused prior components such as long skip connections and simple modifications such as a larger text encoder adapter.

2 Preliminaries

This section introduces terminology for backbone architectures, text and noise conditioning, and open training-data recipes that motivate the paper’s later controlled experiments. It highlights diffusion transformers with flow matching, common conditioning components, and limited disclosure of training data in leading models.

  • Section purpose: The preliminaries establish the terminology and context for the controlled experiments and subsequent modeling and data-design studies.The section focuses on backbone architectures, conditioning mechanisms, and existing open training-data recipes.
  • Text-to-image backbone architectures: Most leading text-to-image systems use diffusion transformers trained with flow matching.The section notes that alternative paradigms exist, but DiTs with flow matching dominate leading systems.
  • Text and noise conditioning mechanisms: Recent diffusion models encode prompts with one or more text encoders and map their features to the diffusion model’s hidden dimension using linear or MLP adapters.The passage identifies both single- and multi-encoder setups and linear or MLP feature adapters.
  • Open text-to-image data recipes: Many leading models release weights publicly without disclosing their training-data recipes.For most models, even dataset sources and scale remain undisclosed, limiting understanding of how to construct strong training data.

3 A Baseline for Controlled Experiments

This section establishes a strong, reproducible baseline for controlled text-to-image diffusion experiments. It fixes the architecture, training and evaluation protocol, and publicly available datasets while varying one design choice at a time before combining improvements into i1.

  • Experimental protocol: Controlled experiments start from the same strong baseline and independently vary one design choice, with successful designs later combined to construct i1.Experiments use the 256-resolution pre-training stage and do not accumulate modifications across experiments.
  • Model: The baseline uses a LightningDiT-XL/2 backbone with cross-attention, QK-norm for training stability, and long skip connections.LightningDiT incorporates RoPE, RMS Norm, and SwiGLU FFN; cross-attention is the default text-injection mechanism.
  • Data: The study exclusively uses publicly available data spanning 7 real-image, 3 synthetic, and 2 text-rendering datasets.These 12 curated image datasets support the controlled experiments.
  • Training and inference: 500K iterations train each baseline with flow matching, batch size 512, and learning rate 1e-4; sampling uses 250-step Euler integration with guidance scale 12.The baseline training covers 25% of the final model’s 2M-step 256-resolution pre-training stage.
  • Evaluation: Evaluation uses DPG-Bench, PRISM-Bench, and LongText-Bench, whose VLM-based scores measure prompt following, aesthetics, and text rendering.DPG and PRISM assess fine-grained prompt following, PRISM additionally evaluates image aesthetics, and LongText targets text rendering.

4 Modeling

The modeling study identifies strong text-conditioning and backbone choices for text-to-image diffusion. i1 uses a single T5Gemma-2B encoder with an expressive adapter, omits AdaLN, and adopts long skip connections with a dual-stream backbone.

  • Text encoder: Encoder–decoder T5Gemma models achieve the best overall performance, outperforming FG-CLIP 2 and decoder-only LLMs and VLMs; i1 therefore uses T5Gemma-2B.Instruction tuning has minimal impact, and larger models do not necessarily perform better.
  • Text encoder combination: Combining T5Gemma-2B with T5Gemma2-1B or FG-CLIP 2 performs best, whereas adding both provides no substantial further gains.Using two separate adapters for repeated T5Gemma-2B embeddings yields a similar improvement, while a shared adapter does not.
  • Text encoder adapter: Increasing adapter capacity for one strong text encoder can capture the gains from multiple encoders without increasing text sequence length, memory, or compute cost.Larger adapters consistently improve performance, while gains beyond two transformer blocks are marginal.
  • Conditioning: AdaLN conditioning on pooled text embeddings, timestep embeddings, or both provides only marginal benefit despite its large parameter cost, so i1 omits AdaLN.With a small adapter, pooled-text AdaLN changes DPG from 84.99 to 82.58; with a large adapter, it changes 85.13 to 84.71.
  • Backbone architecture: Long skip connections improve the performance–parameter trade-off, while the dual-stream backbone has the best trade-off among the evaluated backbone families; i1 uses both.The comparison includes cross-attention, single-stream, and dual-stream backbone families.

5 Data

The section shows that synthetic-caption quality and length strongly affect text-to-image performance: long captions strengthen models overall but require inference-time prompt rewriting for short prompts. It also finds that equal dataset weighting is effective, while diverse data permits substantial image subsampling with only marginal degradation.

  • Synthetic captioning: The synthetic VLM captioner substantially affects downstream text-to-image performance, motivating Qwen3-VL-30B-A3B captions for training i1.The study compares five VLM captioners and selects Qwen3-VL-30B-A3B because its captions produce strong downstream performance.
  • Caption length and prompt rewrite: 0.73 GenEval is achieved by training on long captions and evaluating rewritten prompts, outperforming short-caption training with original, repeated, or rewritten prompts.Long-caption training is stronger overall, but short prompts underperform unless inference-time rewriting expands them.
  • Dataset composition: LongText scores are low for every dataset except TextAtlas, indicating that text rendering relies on specialized text-rich datasets.ImageNet-22K and YFCC achieve the best overall performance, while iNaturalist performs substantially worse, likely because of its narrow domain.
  • Dataset mixing: Equal weighting across datasets is a simple and effective mixing strategy, and removing iNaturalist provides a clear gain across all benchmarks.The selected datasets are equally weighted at each training stage of i1; further removal of additional real datasets offers no substantial improvement.
  • Data magnitude: Subsampling ImageNet-22K from 13.7M to 0.4M images causes only marginal degradation, whereas shrinking to 0.1M produces a substantial drop.With a diverse dataset mix, repeating training data and using fewer unique images can cause only marginal performance degradation.

6 i1-3B: State-of-the-Art Performance Among Fully Open Models

i1 is a 3B-parameter text-to-image model built from the study’s modeling and data findings, with extended low-resolution training and subsequent high-resolution training. It achieves state-of-the-art performance among fully open models on four of five benchmarks and surpasses several leading weight-only models.

  • Data: Equal sampling weights are assigned across six real-image, three synthetic, and two text-rendering datasets, with multiple long synthetic captions generated per image.Caption counts vary by dataset because of resource constraints, ranging from one to five synthetic captions per image.
  • Training: 2M training steps at 256 resolution bring benchmark performance to a plateau, accompanied by improved image quality and text-rendering capability.Performance stabilizes around 500K iterations and largely converges by 2M iterations.
  • High-resolution training: 0.75 →0.92: 512-resolution training substantially improves i1’s LongText score.The model is trained for 0.5M steps at 512 resolution and 0.3M steps at 1024 resolution.
  • Results: State-of-the-art among fully open models: i1 achieves this on all five benchmarks except GenEval and also outperforms several leading weight-only models.The evaluated benchmarks are GenEval, DPG-Bench, PRISM-Bench, CVTG-2K, and LongText-Bench; comparisons include Lumina-Image 2.0, HiDream-I1, and FLUX.1 [Dev].

7 Discussion and Conclusion · Appendix · A Implementation Details

The discussion argues that fully open recipes enable cumulative, attributable research and that strong text-to-image performance can arise from careful, publicly grounded design rather than increasingly sophisticated systems. It also identifies limitations in automated evaluation, scale, visual quality, and exploration scope, while the appendix and implementation details provide supporting material and configuration information.

  • 7 Discussion and Conclusion: Fully open recipes make text-to-image progress more attributable by releasing the model, code, data recipe, and ablations as inspectable artifacts.
  • 7 Discussion and Conclusion: Strong performance can be achieved using moderately scaled 4.4M publicly available datasets and careful exploration of the modeling design space.
  • 7 Discussion and Conclusion: Automated benchmarks primarily emphasize prompt following rather than human preference, limiting how fully evaluation reflects perceived image quality.
  • 7 Discussion and Conclusion: i1 approaches leading weight-only models such as Qwen-Image on the benchmarks, but its generated images remain noticeably inferior in overall visual quality.Failure cases are presented in Appendix B.5.
  • Appendix: Appendix B.5 presents failure cases documenting i1’s remaining visual-quality shortcomings relative to leading weight-only models.
  • 7 Discussion and Conclusion: Resource constraints limited all experiments to models of roughly 3B parameters or smaller, leaving substantially larger-scale validation for future work.
  • A Implementation Details: The modeling and training configurations are documented in the implementation-details section.

A.1 Configuration

The experiments use XL/2-style diffusion configurations with a 29-layer transformer and mixed-precision parameter storage. Controlled models train for 500K iterations, while i1’s high-resolution stages show larger gains at 512 resolution and consistent benefits from timestep shifting at 1024 resolution.

  • General configuration: The default baseline follows XL/2 configurations with hidden size 1152, 16 attention heads, MLP ratio 4.0, patch size 2, and 29 layers.The text encoder uses bf16, while all other parameters use fp32 by default during training and inference.
  • 256-resolution controlled experiments: 500K iterations define all 256-resolution controlled experiments, with cross-attention baselines taking 31.0 hours on a TPU v6e-64 machine.Training time varies across experiments because they use different model components.
  • i1 training: 512-resolution training substantially improves PRISM and LongText, whereas 1024-resolution training has a smaller effect and remains close to its initialization checkpoint.These trends are reported for the high-resolution training stages of i1.
  • i1 training: Applying a training timestep shift of 3.33 consistently improves performance in the 1024-resolution comparison.The comparison is between models trained with a timestep shift and models trained without one.

A.2 Baseline Architectures … B.3 Meta-Prompt for Prompt Rewrite

The appendix defines the controlled-experiment baselines, text-encoder implementation details, qualitative comparisons, inference-setting checks, and a two-mode meta-prompt for rewriting prompts. Together, these sections specify the model variants and procedures used to evaluate and improve text-to-image generation.

  • A.2 Baseline Architectures: Controlled experiments use a fixed baseline architecture, varying one design choice at a time while also validating some choices on single-stream and dual-stream backbones.The cross-attention backbone is the default baseline; the appendix illustrates all three backbone families.
  • A.2 Baseline Architectures: Cross-attention passes text conditioning through inserted cross-attention layers, whereas single-stream concatenates text and noisy-image features under shared weights.The dual-stream variant also concatenates both feature types but uses separate parameters for text and image tokens.
  • A.3 Details on Text Encoders: Text-encoder details select UL2 T5Gemma variants, a 2B decoder for T5Gemma-9B, and FG-CLIP 2 long mode, with right truncation at 256 tokens except 196 for FG-CLIP 2.Encoder-decoder models use final encoder-layer hidden states, while decoder-only models use final-transformer-layer hidden states.
  • B.1 Qualitative Comparison with Stable Diffusion 3 Medium: Qualitative evaluation supplies four additional curated i1 generations and compares them with Stable Diffusion 3 Medium using identical prompts.These comparisons appear in Figures 30 and 31.
  • B.2 Evaluating Other Models under Different Inference Settings: i1 inference uses CFG scale 12 and custom prompt rewriting, exceeding the cited defaults of PixArt-α 4.5, Lumina-Image 2.0 4, SANA 4.5, and Stable Diffusion 3 7.The appendix evaluates whether higher CFG scales and rewritten prompts substantially affect other models.
  • B.2 Evaluating Other Models under Different Inference Settings: Neither higher CFG scales nor rewritten prompts substantially impacts the evaluated other models’ performance.Table 11 compares these settings and marks the default inference configuration with a gray-shaded row.
  • B.3 Meta-Prompt for Prompt Rewrite: Prompt rewriting mitigates the mismatch between long-caption training and short inference prompts, making long-caption training preferable even when original prompts are short.The comprehensive meta-prompt applies different guidelines based on input complexity and includes 20 hand-crafted original–rewritten prompt pairs as in-context examples.
  • B.3 Meta-Prompt for Prompt Rewrite: The meta-prompt classifies inputs into sparse Mode A or dense Mode B and applies mode-specific rewriting protocols targeting 75-150-word, objective, visually grounded outputs.Mode A emphasizes isolation, spatial anchoring, counting, and attribute binding; Mode B prioritizes subjects, exact text transcription, layout, and legibility.

B.4 Ablation of Inference Steps … C.1 Positional Embedding and Normalization

The appendices show that i1 retains strong generation quality with fewer inference steps, while revealing failure cases and benchmark prompt-length differences. Additional modeling experiments motivate combined positional embeddings, sandwich normalization, and shared normalization across text and image streams.

  • B.4 Ablation of Inference Steps: 20 inference steps reduce sampling cost without substantially hurting i1’s generation quality.The model was evaluated with 5, 10, 20, and 50 steps, with 20 identified as the lowest setting without substantial quality loss.
  • B.4 Ablation of Inference Steps: 10 inference steps still produce plausible images, although visual quality degrades gracefully as steps decrease.
  • B.5 Failure Cases of i1: i1 sometimes generates low-fidelity faces, unnatural expressions, and malformed hands or limbs in scenes containing multiple small human figures.The model also sometimes violates physical properties, producing physically implausible images such as incorrect mirror reflections.
  • B.6 Prompt Length Distributions of Benchmarks: GenEval has substantially shorter prompts than the other four benchmarks, motivating analyses of short-prompt performance and inference-time prompt enhancement.
  • C Additional Results on Modeling Designs: Appendices C.1 and C.2 provide additional experiments motivating the positional embedding, normalization, and VAE choices used in the final i1 model.
  • C.1 Positional Embedding and Normalization: Combining sinusoidal and RoPE positional embeddings improves benchmark performance over using either embedding alone.The improvement applies across cross-attention and dual-stream backbone families.
  • C.1 Positional Embedding and Normalization: Sandwich normalization stably improves performance over standard pre-norm across backbone architectures and benchmarks.Sandwich normalization normalizes both the inputs and outputs of attention and feed-forward modules.
  • C.1 Positional Embedding and Normalization: Sharing normalization parameters across text and image modalities consistently improves MMDiT performance over modality-specific normalization.

C.2 VAEs … C.5 Comparing Backbone Families with Training FLOPs

The appendix finds that FLUX.2 VAE offers the most balanced benchmark performance, while VA-VAE excels on DPG-Bench but has weaker reconstruction and LongText results. Alternative text-encoder settings preserve prior trends, system prompts provide a minor boost, and dual-stream backbones perform best across training FLOPs.

  • C.2 VAEs: FLUX.2 VAE achieves the most balanced performance across all benchmarks, while VA-VAE is strongest on DPG-Bench but performs much worse on LongText.The passage attributes VA-VAE’s DPG-Bench strength to alignment with pretrained semantic features.
  • C.2 VAEs: VA-VAE introduces noticeable character distortions and corruptions in text-rich images, unlike FLUX.2 VAE and Qwen-Image VAE.The weaker reconstruction capability may explain VA-VAE’s lower LongText performance.
  • C.2 VAEs: 27.96 PSNR and 0.79 SSIM are VA-VAE’s ImageNet reconstruction scores, below FLUX.2 VAE’s 31.46 and 0.90 and Qwen-Image VAE’s 33.42 and 0.92.VA-VAE also introduces visible corruption and distortion in text-rich reconstructions, possibly because it was trained on ImageNet, which lacks text-rich images.
  • C.3 Comparing Text Encoders under Alternative Settings: Using a larger adapter preserves the text-encoder performance trends observed with the smaller MLP adapter across benchmarks.The larger adapter consists of two transformer blocks, as in the final i1 recipe.
  • C.3 Comparing Text Encoders under Alternative Settings: Removing AdaLN likewise produces text-encoder trends highly similar to those in the default setting across benchmarks.In the baseline, AdaLN injects pooled text embeddings together with timestep embeddings into the backbone.
  • C.4 Applying System Prompts to Text Encoders: A system prompt wrapping the text-to-image prompt brings a minor performance boost for Qwen3-VL text encoders.The method discards hidden states for the system prefix and retains only those corresponding to the image prompt.
  • C.5 Comparing Backbone Families with Training FLOPs: Dual-stream backbones achieve the best overall performance when comparing cross-attention, single-stream, and dual-stream models across estimated training FLOPs.Training FLOPs are computed with JAX/XLA’s cost_analysis for one forward and backward pass.

C.6 Validating Modeling Designs on Larger Models

Validation across dual-stream MMDiT model sizes confirms that larger text encoder adapters generally improve performance-parameter trade-offs, while several other modeling trends are not consistent across scales. Under training-FLOPs comparisons, most findings remain unchanged, except that AdaLN benefits larger transformer adapters.

  • Performance vs. model size: Models were trained at widths 1152, 1296, 1440, 1584, and 1728 to validate modeling findings across dual-stream MMDiT scales.
  • Performance vs. model size: Larger text encoder adapters consistently deliver better performance-parameter trade-offs across model sizes.
  • Performance vs. model size: Removing AdaLN helps with MLP text encoder adapters, whereas larger transformer adapters show similar performance with or without AdaLN at comparable parameter counts.These results still suggest that noise conditioning may not be necessary in text-to-image diffusion models.
  • Performance vs. model size: Combining Sinusoidal and RoPE embeddings, sandwich normalization, and shared image-text normalizations does not improve performance consistently across model scales.
  • Performance vs. training FLOPs: 18.9% of parameters belong to AdaLN in the dual-stream baseline, and with larger transformer adapters AdaLN achieves a better performance-FLOPs trade-off than omitting it.Most trends remain unchanged when performance is compared against estimated training FLOPs for trainable modules.

C.7 Validating Long Skip Connection Results on Other Backbones … D.1 Additional Designs in Synthetic Captioning

The appendices validate long skip connections and text-feature adapters across architectures, explore skip-connection variants, and provide additional evidence for synthetic-captioning design choices. Results favor long skips, larger text adapters, the default symmetric skip pattern, center-cropped captioning inputs, and multiple captions per image.

  • C.7 Validating Long Skip Connection Results on Other Backbones: Removing long skip connections noticeably reduces performance across most backbones and benchmarks, especially on DPG and LongText.The validation uses XL/2-sized baseline variants trained with and without long skip connections.
  • C.8 Text Feature Adapter vs. Image Feature Adapter: Transformer adapters for text features substantially improve performance across backbones, whereas the same adapters on image features yield much smaller gains.This indicates the benefit is specific to adapting pre-trained text features rather than merely increasing parameter count.
  • C.9 Exploring Variants of Long Skip Connections: The default symmetric long skip pattern performs best overall, although the (i + 21) variant slightly improves DPG on the cross-attention backbone.The comparison tests connections from layer i to layers (i + 14) and (i + 21).
  • C.9 Exploring Variants of Long Skip Connections: None of the tested long skip layer ranges consistently outperforms the default configuration, in which layers 1-14 have skip connections.The ablation varies the layer interval receiving long skip connections.
  • D Additional Results on Data Designs: Appendix D.1 adds controlled experiments supporting center-cropping before captioning and generating multiple captions per image.Appendix D.2 separately provides additional results supporting the conclusion on data mixing.
  • D.1 Additional Designs in Synthetic Captioning: Captions generated from cropped and resized images achieve similar downstream performance to captions from full images, while preprocessing improves captioning speed.Center-cropping also avoids captions describing objects later removed from the training images.
  • D.1 Additional Designs in Synthetic Captioning: Generating five captions per image produces modest improvements on PRISM and LongText, with larger benefits when unique training images are limited.Multiple captions provide an additional axis of data scaling beyond increasing the number of images.

D.2 Experiments under Equal Dataset Weights … E.3 Caption Length Distribution for Different VLM Captioners

The supplementary sections examine dataset weighting, training-data composition and resolution filtering, and synthetic-captioning choices. They show that single-dataset upsampling is not consistently beneficial, document the datasets and resolution criteria used, and relate caption length to text-to-image performance.

  • D.2 Experiments under Equal Dataset Weights: ×3 upsampling of any single dataset does not consistently improve performance across all benchmarks.The section further evaluates ×5 upsampling and reports that its results are consistent with the earlier finding.
  • E.1 Visualizations of Images: Random samples of images from each training dataset are provided as visualizations.The figure documents the image contents associated with the training datasets.
  • E.1 Visualizations of Images: The training study uses 12 publicly available image datasets, including 7 real-image and 3 synthetic datasets.The listed real-image datasets include ImageNet-22K, YFCC100M, RedCaps, Megalith, Pexels, iNaturalist 2024, and Places365-Challenge 2016.
  • E.2 Image Resolution Statistics: For 256-resolution training, i1 uses all image datasets except iNaturalist.For 512/1024-resolution training, images are filtered by shorter-edge size and datasets with fewer than 0.3M remaining images are removed.
  • E.2 Image Resolution Statistics: At 512 resolution, training uses all datasets except YFCC and iNaturalist, while 1024-resolution training uses FLUX-Reason, TextAtlas, RedCaps, GPT-Edit, and Midjourney v6.RenderedText is excluded at 1024 resolution despite satisfying the image-count requirement because of its low quality.
  • E.2 Image Resolution Statistics: The RedCaps dataset contains 5M images rather than the 12M reported in its original paper because many URL-provided images are inaccessible.The discrepancy reflects image availability in the dataset used here.
  • E.3 Caption Length Distribution for Different VLM Captioners: The VLM used for synthetic captioning substantially affects downstream text-to-image performance.ImageNet-22K captions were generated with Qwen2-VL 2B, Qwen2.5-VL 3B, Qwen3-VL-2B, Qwen3-VL-4B, or Qwen3-VL-30B-A3B using the same meta-prompt.
  • E.3 Caption Length Distribution for Different VLM Captioners: Captioner prompt sets associated with strong text-to-image performance also tend to have longer overall token-sequence-length distributions.Caption lengths are measured with the T5Gemma tokenizer.

E.4 Caption Length Distribution for Different Datasets · E.5 Meta-Prompt for Synthetic Captioning · E.6 Sanity Check for Data Leakage

The appendix compares synthetic-caption length distributions, specifies dataset-specific captioning prompts, and checks potential evaluation-train caption overlap. Caption lengths vary across datasets without substantial differences, while text-rendering data receive prompts that incorporate ground-truth text details and FLUX-Reason undergoes an exact-overlap check against PRISM-Bench.

  • E.4 Caption Length Distribution for Different Datasets: 10K random images per dataset are used to compare synthetic-caption sequence lengths with the T5Gemma tokenizer.The distributions are plotted in Figure 49.
  • E.4 Caption Length Distribution for Different Datasets: Caption length distributions vary across datasets, but no dataset differs substantially from the others.
  • E.5 Meta-Prompt for Synthetic Captioning: Most image datasets use the minimal synthetic-captioning prompt, “Describe the image in detail using one paragraph.”
  • E.5 Meta-Prompt for Synthetic Captioning: Text-rendering datasets include ground-truth text annotations in the prompt to reduce hallucinations in VLM-generated captions.
  • E.5 Meta-Prompt for Synthetic Captioning: TextAtlas prompts provide the ground-truth annotation as reference alongside a request for a detailed one-paragraph description.
  • E.5 Meta-Prompt for Synthetic Captioning: RenderedText prompts expose the number and comma-separated content of text lines and request transcription, font, size, color, location, and rotation angle.RenderedText contains handwritten text rendered on digitally modeled 3D paper under varied font and lighting conditions.
  • E.6 Sanity Check for Data Leakage: For datasets other than FLUX-Reason, synthetic VLM captions make overlap with evaluation benchmark prompts unlikely.
  • E.6 Sanity Check for Data Leakage: FLUX-Reason combines its original “caption_detail” captions with four synthetic caption sets, and the authors perform exact string matching against PRISM-Bench for overlap.Half of PRISM-Bench’s prompts were reportedly selected from FLUX-Reason and then removed from that dataset.
Loading 2606.11289v1…