Lens: Rethinking Training Efficiency for Foundational Text-to-Image Models

Introduction

• 19.3% of Z-Image’s training compute is used by 3.8B-parameter Lens, which matches or surpasses larger state-of-the-art models across multiple benchmarks.The comparison reports 192K A100 GPU hours for Lens and 314K H800 GPU hours for Z-Image, normalized using peak BF16 TFLOPS.Source: §1 · p. 2, Figure 2 · p. 3

Data Information Density per Training Batch

• Lens trains on 800M GPT-4.1-captioned image-text pairs, with captions averaging 109 words, to provide richer semantic supervision than short captions.Each batch mixes resolutions of 512^2, 768^2, and 1024^2 with diverse aspect ratios; this supports generalization to unseen ratios and resolutions up to 1440^2 without costly high-resolution training.Source: §1 · p. 3

Convergence Speed

• Lens studies VAE variants directly in the T2I pipeline and adopts GPT-OSS, whose stronger language conditioning accelerates convergence and supports multilingual inputs despite English-only training.The VAE evaluation uses a 130M subset of the training data; GPT-OSS enables generalization to languages including Chinese and French.Source: §1 · pp. 3–4

Reinforcement Learning Post-Training

• Lens-RL-8K uses taxonomy-driven prompt coverage to improve generation across scenarios, suppress artifacts, and avoid degradation on input types omitted by narrow RL data.The paper emphasizes that RL data should be diverse and cover the original training distribution.Source: §1 · p. 4

Reasoner Module

• A reasoner rewrites ambiguous requests into detailed prompts, while training-free system-prompt search tunes its guidelines to better align prompts with Lens.Experiments report results both with and without the reasoner for fair comparison.Source: §1 · p. 4

Inference Efficiency

• Lens generates a 1024^2 image in 3.15 seconds on one NVIDIA H100 using 20 denoising steps; its 4-step distilled Lens-Turbo takes 0.84 seconds.The smaller model also enables faster inference than larger models at the same number of denoising steps.Source: §1 · pp. 2, 4

2.1 Pre-training Data: Lens-800M

• Dense-caption training outperforms brief and mixed-caption Lens-Toy variants on GenEval, improving data utilization efficiency.Lens-800M contains approximately 800M images from four complementary sources, captioned in detail by GPT-4.1; image text is preserved in its original language. The captions use a multi-stage data-cleaning pipeline.Source: §2.1 · pp. 5–6, Figure 4 · p. 5

2.2 Architecture

• FLUX.2’s semantic VAE achieves the best Lens-Toy generation performance and accelerates convergence, so Lens adopts it.The paper evaluates VAEs directly in text-to-image generation rather than using reconstruction fidelity or class-conditional ImageNet generation as proxies.Source: Figure 5 · p. 5, §2.2 · p. 6
• GPT-OSS is selected after English-only Lens-Toy comparisons; strong encoders aid convergence and multilingual inputs, while a swappable Reasoner refines prompts.GPT-OSS can serve as the Reasoner without extra GPU memory cost; the default Reasoner is GPT-5.5.Source: §2.2 · p. 7, Figure 7 · p. 8, Figure 8 · p. 8

2.3 Pre-training

• Mixed-resolution pretraining across 27 buckets yields generalization to unseen sizes, aspect ratios from 1:2 to 2:1, and image areas up to 1440^2.Training uses 400K fixed-resolution 512 × 512 iterations followed by 400K mixed-resolution iterations; resolution generalization reduces the need for costly high-resolution training.Source: §2.3 · pp. 8–9

2.4 Post-training

• Lens-RL-8K’s broad taxonomy-driven prompt coverage and prompt-specific rubrics support improvements across diverse scenarios, while reduced or biased coverage limits generalization.The dataset contains 8,406 prompts spanning ten broad categories; GPT-4.1 generates ten sample-aware rubrics per prompt, supplemented by a global coherence and plausibility rubric.Source: §2.4 · pp. 9–10
• Lens-Turbo distills Lens-RL into a 4-step generator that largely preserves image quality and prompt-following ability.Distillation combines DMD2, decoupled-DMD, SenseFlow, and R1 regularization on the adversarial loss for stability.Source: §2.4 · p. 10

2.5 Inference

• Inference uses a Reasoner followed by 20-step Lens generation with CFG 5.0, while training-free iterative system-prompt search refines the Reasoner’s prompts.The search uses GPT-5.5 to rewrite the prompt based on prior prompts and textual summaries of generation failures, without model training.Source: §2.5 · p. 10

3 Comparison with State-of-the-art Models

• OneIG (EN) [6] evaluates 1,120 prompts across general objects, portraits, anime and stylization, text rendering, and knowledge and reasoning; GenEval [7] tests object-centric compositional alignment across six templated tasks.OneIG reports dimension-specific scores including subject alignment, text accuracy, reasoning, style, diversity, and overall performance; GenEval uses detector- and classifier-based verification.Source: §3 · pp. 10–11
• Table 2 compares Lens at 20-step inference and Lens-Turbo at 4-step inference across four benchmarks, reporting overall scores for OneIG, GenEval, and LongText and three CVTG metrics.For CVTG, the reported metrics are average score, normalized edit distance (NED), and CLIP score; detailed comparisons appear in Appendix B.1.Source: Table 2 · p. 11
• LongText (EN) [31] tests text rendering in 160 prompts spanning eight text-rich scenarios, with short and longer text passages.Prompts include signboards, labeled objects, printed materials, webpages, slides, posters, captions, and dialogues, with text lengths of roughly 10–30 or 30–50 words.Source: §3 · p. 11
• CVTG [32] evaluates complex visual text generation across 2–5 text regions, specifying content, position, length, and attributes such as color, font, and size.Source: §3 · p. 11

• The visualizations show Lens-Turbo generating high-quality images, realistic portraits, multilingual text, and following user instructions written in different languages.Examples cover general image generation (Figures 9–10), portraits (Figures 11–12), multilingual visual text (Figures 13–14), and multilingual prompt following (Figures 15–16); all images have an area of 1440^2 pixels and varying aspect ratios.Source: §4 · pp. 11–12

5 Conclusion

• Lens, a 3.8B-parameter foundational T2I model, is competitive with and sometimes surpasses larger state-of-the-art models at substantially lower training cost.The conclusion also reports fast inference and argues that efficient training strategies can improve foundational T2I model scalability.Source: §5 · p. 12
• Dense captions and mixed-resolution, mixed-aspect-ratio pre-training on Lens-800M increase data information density, while selected VAE and language encoder designs accelerate convergence.These strategies are presented as the basis for Lens’s strong generation quality at lower training cost.Source: §5 · p. 12
• RL post-training on Lens-RL-8K, a reasoner, training-free system-prompt search, and few-step distillation further enhance Lens.These optimizations target model quality and inference efficiency.Source: §5 · p. 12

A.1 Foundational T2I Models

• Latent diffusion models such as Stable Diffusion established compressed-latent generation to reduce costs, while SDXL adds high-resolution refinements including stronger conditioning and multi-aspect-ratio training.The cited works are Stable Diffusion [33] and SDXL [34].Source: §A.1 · p. 27
• Recent systems including FLUX, SANA, HiDream-I1 [38], Qwen-Image [4], Hunyuan-Image-3.0 [5], and Z-Image [1] advance visual quality, prompt following, and inference efficiency.The related-work discussion also identifies progress in text rendering, image editing, and reference consistency.Source: §A.1 · p. 27
• Stable Diffusion 3 and 3.5 use MMDiT-style architectures with rectified-flow training [9], reflecting a broader shift toward scalable diffusion Transformers [36].This passage situates the shift within the recent evolution of T2I backbones and objectives.Source: §A.1 · p. 27
• Janus-Pro [43], Transfusion [44], and BAGEL [45] integrate image generation with language modeling or visual understanding, but such systems often demand large models, massive data, and complex designs.This trend motivates pursuing competitive generation quality with substantially reduced computational cost.Source: §A.1 · p. 27

A.2 Post-training for T2I Models

• T2I post-training has expanded from supervised fine-tuning to preference optimization and RL, directly targeting human preferences, reward judgments, or task-specific criteria.These approaches aim to improve model behavior beyond likelihood-based pre-training.Source: §A.2 · p. 27
• DPO-style diffusion methods optimize positive-negative image pairs or preference sets; Diffusion-DPO [46], D3PO [47], SPO [48], and variants [49] [50] [51] offer relatively simple, stable alternatives to online RL.They optimize preference objectives without requiring explicit reward maximization during sampling.Source: §A.2 · pp. 27–28
• Policy-gradient RL treats generation as sequential decision-making; Flow-GRPO [53] and variants [54] [55] optimize flow trajectories, while DiffusionNFT [24] and AWM [56] explore related objectives.Flow-GRPO converts deterministic sampling into an equivalent stochastic formulation; related frameworks use online optimization and advantage-aware or negative-aware objectives.Source: §A.2 · p. 28
• Structured rewards for RL should assess aesthetics, prompt faithfulness, object correctness, composition, text rendering, and safety to reduce reward hacking and misalignment.This motivates Lens’s rubric-based post-training strategy, which provides explicit, multidimensional criteria for RL optimization.Source: §A.2 · p. 28

A.3 Distillation for T2I Models

• Iterative diffusion and flow sampling is expensive, and training-free samplers such as DDIM [60], DPM-Solver [61], and UniPC [62] can lose quality at very low budgets.Degradation is especially notable for high-resolution text-conditioned synthesis under strong classifier-free guidance.Source: §A.3 · p. 29
• Trajectory- and consistency-based distillation enables few-step generation, but quality can still degrade under very small step budgets or complex text conditioning.Examples include progressive distillation [63], consistency models [64], Latent Consistency Models [65], and InstaFlow [66], which straightens probability-flow trajectories and improves noise-data couplings.Source: §A.3 · p. 29
• Distribution-level and adversarial supervision advances few-step distillation, yet results remain sensitive to teacher guidance, timestep design, fake-score estimation, and adversarial stability.Adversarial Diffusion Distillation [67] combines score distillation with an adversarial objective; Distribution Matching Distillation [68] matches student and target distributions, with variants [69], [27], [70], and [28] improving stability, guidance distillation, or alignment.Source: §A.3 · pp. 29–30

A.4 VAE

• Reconstruction-optimized latents can be semantically entangled and difficult for diffusion models to learn, slowing convergence and potentially harming text-to-image generation and editing.The cited work argues that excessive reconstruction pressure favors low-level details over semantic structure [73].Source: §A.4 · p. 30
• REPA-E, Unified Latents [75], and Latent Forcing [76] align tokenizer training or latent trajectories more closely with generative learning.The approaches respectively align encoder representations with diffusion Transformer features, jointly optimize reconstruction and generation, and impose latent-level constraints while reorganizing the diffusion trajectory.Source: §A.4 · pp. 30–31
• VTP [10], MagViT-v2 [77], VAR [78], TiTok [79], and CLIP- or DINO-based distillation explore richer semantic, structured, or learnable latents.Together, these approaches reflect a shift from reconstruction-optimal tokenizers toward generation-oriented designs.Source: §A.4 · p. 31

B.1 Detailed Benchmark Results

• Tables 3–5 compare Lens with state-of-the-art models on OneIG (EN), GenEval, LongText (EN), and CVTG.The supplied passage identifies the comparison scope but does not provide table values or winners.Source: §B.1 · p. 31

B.2 Various Reasoners

• Table 6 compares Lens variants using no reasoner, GPT-5.5, GPT-OSS-20BA3B, and Qwen3 reasoners at 0.6B, 1.7B, and 4B sizes.The passage describes the variants but supplies no comparative results.Source: §B.2 · p. 31

C.1 Prompt-rubric Visualization

• A playground example specifies one bright red kickball, five empty swings, and one school-age child, with rubrics checking those exact counts.The prompt and rubrics also specify where the child and swings should appear.Source: §C.1 · pp. 31–32
• The playground rubrics check the child’s left-foreground placement, the swings behind, and the action of gripping the kickball at waist level.These criteria translate spatial relationships and an action in the prompt into explicit checks.Source: §C.1 · pp. 31–32
• Other playground checks require the fence sign to match “PLAYGROUND RULES – WALK PLEASE” and verify the kickball, sneakers, knees, and midday shadows.The rubric set includes both OCR alignment and attribute-accuracy checks.Source: §C.1 · pp. 31–32
• A structural-integrity rubric checks that the playground image is coherent and physically plausible, without missing, duplicated, disconnected, or misplaced elements.It also checks that subjects, objects, perspective relationships, and spatial connections appear complete and naturally integrated.Source: §C.1 · p. 32

• Figure 17 shows Lens-800M training examples as densely captioned image-text pairs.Source: §C.2 · p. 33

D.1 RL Details

• DiffusionNFT trains positive and negative flow-matching velocities so rewards steer generation toward desirable directions and away from undesirable trajectories [24].High-reward samples emphasize the positive velocity; low-reward samples emphasize the negative velocity.Source: §D.1 · pp. 33–34
• DiffusionNFT maps variable raw rewards to bounded optimality probabilities, stabilizing the balance between positive-direction learning and negative-direction avoidance [24].The normalization factor is typically the global reward standard deviation.Source: §D.1 · pp. 34–35
• RL finetuning runs for 180 steps on Lens-RL-8K with 64 A100 GPUs, using mixed-resolution buckets and LoRA.The nine aspect ratios span 736×1472 to 1472×736 at a fixed 1024^2 base area; LoRA uses rank 64 and scaling factor 128. Optimization uses 48 groups of 24, a second-order ODE sampler, and adaptive time weighting; the KL penalty is 1 × 10^-4, and the learning rate is 3 × 10^-4 [24].Source: §D.1 · p. 35

D.2 Distillation Details

• DMD distills a multistep model by matching the student distribution to the data distribution, using frozen-teacher and fake-score gradients [68].The teacher score pulls samples toward the target distribution, while the fake score counteracts student over-concentration.Source: §D.2 · p. 35
• Distillation uses a curated 100K-image Lens-800M subset balanced across major generation scenarios, with mixed-resolution loading for student simulation and discriminator training.Images are selected using aesthetic and other quality-related metrics; captions condition the DMD objectives, and paired images provide real discriminator samples.Source: §D.2 · p. 35
• The RL-aligned student and fake score model are fully optimized with decoupled-DMD guidance and distribution-matching terms, while the teacher remains frozen [27].An adversarial branch [26] compares equally noised real and student samples using frozen-teacher features rather than fake-score features; its practical settings include γ = 1.0, α = 0.1, λd = 0.1, and λg = 0.001.Source: §D.2 · pp. 35–36
• Each distillation step updates the fake score model and discriminator four times before one student update, then applies IDA with µ = 0.03 [26, 28].Training targets 4-step generation with guidance scale 5.0, using 8 A100 GPUs, per-GPU batch size 4, and up to 1K global steps.Source: §D.2 · p. 36

E.1 Prompt for Image Captioning

• Image captions are instructed to describe salient objects, attributes, relationships, background, and visible text, adding relevant world knowledge when recognizable content appears.The prompt asks for a comprehensive, accurate caption that can support visual reconstruction.Source: §E.1 · p. 37

E.2 Prompt for Lens-RL-8K Dataset Construction

• Lens-RL-8K construction generates exactly five diverse, 30–60-word photorealistic prompts per entity, incorporating every required keypoint in each prompt.Keypoints cover scene composition, counts, position, attributes, color, relations, and visible text.Source: §E.2 · p. 37

E.3 Prompt for Rubric Generation

• The rubric-generation prompt requests up to 10 short named criteria, such as object-count consistency or OCR alignment, in a single valid JSON object.The instructions require preserving key names and returning JSON parsable by Python.Source: §E.3 · p. 38

E.4 Prompt for RL Reward Generation

• For RL reward generation, GPT-4.1-mini evaluates each image against a visual-quality rubric and returns 1 for full satisfaction or 0 for clear failure.The evaluation uses the image-generation prompt and the corresponding criterion.Source: §E.4 · p. 38

E.5 Prompt for General Inference

• The general-inference reasoner rewrites user input into one coherent, visually grounded prompt while preserving intent and resolving only direct, unambiguous inferences.It may add consistent visual details, but must not change described content or invent hidden facts.Source: §E.5 · p. 39

E.6 Prompt for GenEval Benchmark

• The GenEval reasoner preserves metadata-specified objects, counts, colors, and spatial relations while making them clear and easy to detect in uncluttered scenes.Its rules also avoid adding unrequested evaluated objects or other details that could cause prompt drift.Source: §E.6 · pp. 40–41

E.7 Prompt for OneIG Benchmark

• The OneIG reasoner rewrites inputs into renderable prompts while preserving explicit benchmark-critical details and avoiding unsupported additions.The supplied instructions specifically call for treating comma-separated anime and tag-style inputs as required visual constraints.Source: §E.7 · p. 42

E.8 Prompt for LongText and CVTG Benchmarks

• For LongTextBench and CVTG, the reasoner rewrites each benchmark prompt as one improved image-generation prompt focused on accurate text rendering and faithful scene meaning.The benchmarks cover signs, labels, printed documents, webpages or app screens, slides, posters, captions, lower-thirds, subtitles, and comic dialogue bubbles.Source: §E.8 · pp. 43–44
• Include every required visible text string exactly as supplied, preserving its characters and order, and enclose each string in single quotation marks.If the original prompt conflicts with the required-text list, the list is authoritative.Source: §E.8 · p. 44
• Use a text-first composition with large, flat, front-facing, high-contrast text regions; simplify decoration rather than shrinking or obscuring required text.Avoid difficult-to-read treatments such as rotated, perspective-skewed, tiny, crowded, reflective, or highly stylized text unless the prompt requires them.Source: §E.8 · p. 44
• Preserve the original category, subjects, setting, style, composition, and spatial relationships without adding unsupported text, entities, events, or factual claims.Keep background details only when they do not compete with the text, and organize already-detailed prompts for readability rather than adding scene elements.Source: §E.8 · p. 44
• For signs, make the main sign large, flat, front-facing, and high-contrast; for labels, place each clearly beside its corresponding object.Use separate signs only when explicitly requested, and preserve each label’s association with the object it describes.Source: §E.8 · p. 44
• For printed materials, show the surface flat or facing the viewer with clear hierarchy; for webpages and app screens, use clean, front-facing layouts with readable interface elements.Long documents should use spacious text blocks, while non-required interface text should remain minimal or abstract.Source: §E.8 · pp. 44–45
• Use simple, readable layouts for slides and posters, separating required content into clear blocks and keeping decorative elements away from the text.Slides may use numbered cards or a two-column grid; posters should preserve the original topic and mood without inventing event details.Source: §E.8 · p. 45
• For captions, subtitles, news lower-thirds, and dialogue, use wide high-contrast text areas and clear reading order; output one concise, text-focused prompt without explanations or alternatives.The reasoner may add concrete layout, typography, lighting, or contrast details when they help readability, but should avoid generic quality boosters and generation parameters.Source: §E.8 · p. 45

F Broader Impacts and Limitations

• Reducing training costs can lower barriers to visual-content research and development across image generation, design assistance, education, and creative applications.Lens aims to make high-quality text-to-image generation more efficient and accessible.Source: F Broader Impacts and Limitations · pp. 45–46
• Lens may be misused to generate misleading, biased, or harmful images; its reasoner rejects inappropriate requests, while responsible deployment also needs moderation, provenance tracking, misuse detection, and bias consideration.Source: F Broader Impacts and Limitations · p. 46
• English-focused training can leave multilingual generation quality and prompt-following below English performance, and cannot replace direct training on diverse multilingual text–image data.Lens can generalize to prompts such as Chinese and French, but its quality and prompt-following may be lower than with English prompts.Source: F Broader Impacts and Limitations · p. 46
• Lens may struggle to render non-English text, including Japanese and French, and may produce artifacts in undercovered scenarios, rare compositions, complex layouts, or challenging concepts.Limited training-data coverage of text patterns contributes to rendering errors; insufficient scenario coverage likely contributes to visual artifacts.Source: F Broader Impacts and Limitations · p. 46
• Future work could expand multilingual and text-rich data, improve long-tail coverage, strengthen post-training or refinement, and develop more robust safety mechanisms for deployment.Source: F Broader Impacts and Limitations · p. 46