How a 3.8B Model Beats 6B+ Models Using Just 20% of the Compute – Inside Microsoft Lens

Small Models Can Compete

Training text‑to‑image (T2I) models traditionally requires tens of thousands of GPU hours, making large‑scale models (6 B, 9 B, 20 B, even 80 B parameters) prohibitive for many teams. Microsoft’s Lens team reduced the parameter count to 3.8 B, cutting per‑step computation and inference latency. Despite the smaller size, Lens matches or exceeds larger models on four major benchmarks: GenEval (0.557, higher than 6 B Z‑Image), LongText (0.930, a new open‑source record), CVTG (NED 0.951, CLIP 0.814), and OneIG.

In terms of compute, Lens used 192 K A100 GPU‑hours, whereas Z‑Image required 314 K H800 GPU‑hours. After normalising to a common peak‑performance baseline, Lens consumed only 19.3 % of the compute of the 6 B baseline.

Information Density Is Key

Lens tackled the low information density of traditional short captions (e.g., "a photo of a cat") by generating dense captions with GPT‑4.1. Each caption averages 109 words, describing objects, attributes, spatial relations, actions, and background. The entire Lens‑800M dataset (800 M images) was re‑annotated this way.

Ablation experiments on a 1.3 B‑image subset compared short, dense, and mixed captioning. Dense captions consistently outperformed the other two on GenEval, demonstrating that a single dense caption provides a stronger training signal than a short one.

On the image side, Lens introduced mixed‑resolution and aspect‑ratio training. Each batch contains images at 512, 768, and 1024 pixels with nine aspect ratios (1:2 to 2:1), forming 27 resolution buckets. This exposes the model to diverse scales and compositions, boosting visual information density and enabling zero‑shot generalisation to unseen resolutions up to 1440 px.

Data cleaning involved a nine‑step pipeline: corrupted file removal, resolution filter (<384 px²), NSFW filtering, aesthetic score (<3) removal, watermark detection, clarity filter, entropy filter, brightness filter, and near‑duplicate removal using CLIP embeddings (cosine similarity > 0.985).

The final dataset mixes four sources: 4.558 B real images (57 %) and 3.444 B synthetic images (43 %).

Convergence Speed Matters

Lens evaluated four VAEs and found the FLUX.2 semantic VAE superior on GenEval, citing tighter latent spaces and richer semantics that ease text‑image alignment.

Four language encoders were compared: GPT‑OSS (20 B MoE), Qwen3‑0.6B, Qwen3‑1.7B, and Qwen3‑4B. GPT‑OSS achieved the best English performance and, surprisingly, enabled zero‑shot multilingual generation (Chinese, French, Japanese, Spanish) despite training only on English data, vastly outperforming the Qwen3 variants.

The overall architecture follows an MMDiT (multimodal diffusion Transformer) style with 48 blocks. Each block contains separate image and text branches that perform self‑attention before cross‑modal interaction. The image branch uses RoPE positional encoding, aiding generalisation to unseen resolutions. Text features are extracted from GPT‑OSS layers 4, 12, 18, 24, concatenated, and linearly projected.

Post‑Training Enhancements

After pre‑training, Lens‑Base still produced occasional artifacts (structural errors, blurry details, physical inconsistencies). Reinforcement‑learning (RL) fine‑tuning was applied using the Lens‑RL‑8K dataset (8 406 prompts covering ten high‑level categories and numerous sub‑categories). Each prompt was expanded with 5 descriptive dimensions (attributes, spatial relations, quantity, interaction, colour) via GPT‑4.1.

Evaluation rubrics (10 per prompt) were also generated by GPT‑4.1, plus a global rule ensuring structural coherence and physical plausibility. DiffusionNFT, with GPT‑4.1‑mini as the reward model, sampled 48 prompt‑rubric pairs per step, generated 24 images at varying resolutions, and updated the policy over 180 steps on 64 A100 GPUs.

Ablation showed data coverage matters: using ¼ of the RL dataset yielded GenEval 0.916, ½ gave 0.920, and the full set achieved 0.930. Removing text‑related prompts degraded CVTG and OneIG scores (NED from 0.951 to 0.928, CLIP from 0.814 to 0.795), confirming the importance of textual cues.

The “Reasoner” module expands vague user prompts into detailed, training‑distribution‑aligned prompts. By default it uses GPT‑5.5, but any LLM (e.g., GPT‑OSS) can replace it without extra memory cost.

Finally, Lens‑Turbo distills the RL‑fine‑tuned model using a blend of DMD2, decoupled‑DMD, and SenseFlow techniques plus R1 regularisation. The distilled model generates images in four diffusion steps (no CFG) at 0.84 s for a 1024 px image, while preserving quality and prompt fidelity.

Inference defaults to 20 diffusion steps with CFG 5.0, supports arbitrary aspect ratios from 1:2 to 2:1, resolutions up to 1440 px, and multilingual prompts (English, Chinese, French, Japanese, Spanish).

In summary, by systematically rethinking training efficiency across model size, data density, resolution strategy, architecture choices, and post‑training optimisation, the 3.8 B‑parameter Lens model achieves higher quality, faster inference, and dramatically reduced compute cost compared to traditional 6 B‑plus T2I models.