Lens-3.8B-bf16 (MLX)

Apple MLX conversion of the denoising transformer (DiT) from microsoft/Lens — a 3.8B-parameter foundational text-to-image model — for fast inference on Apple Silicon. bf16, full precision.

This repo contains the DiT only (MIT-licensed). The full pipeline also uses the GPT-OSS-20B text encoder (Apache-2.0) and the FLUX.2 semantic VAE, which the loader pulls from their own sources rather than re-hosting here (see License below).

component	parity vs PyTorch reference
GPT-OSS text features	per-layer cosine ≈ 0.998
Lens DiT (this repo)	cosine 0.999999
FLUX.2 VAE decode	PSNR 57.65 dB
full end-to-end image	PSNR 45.26 dB

Generates a 1024×1024 image in ~33 s on Apple Silicon (20 steps, ~39 GB peak).

Usage

from lens_mlx.pipeline_mlx import LensPipeline   # github.com/xocialize/lens-mlx

# `base` = a microsoft/Lens snapshot providing the tokenizer, GPT-OSS encoder, and FLUX.2 VAE.
pipe = LensPipeline.from_pretrained(base, dit_repo="mlx-community/Lens-3.8B-bf16")
img = pipe("A serene lake below snow-capped mountains, golden hour.",
           height=1024, width=1024, num_inference_steps=20, seed=42)
img.save("out.png")

Conversion

Converted from microsoft/Lens with recipes/convert_lens.py (lens-mlx). The DiT is pure Linear + RMSNorm; weights map 1:1 (no transpose) and every tensor is materialized before save. Layer-by-layer parity against the PyTorch reference is in the lens-mlx test suite.

License

DiT weights (this repo): MIT, inherited from microsoft/Lens.
GPT-OSS-20B encoder: Apache-2.0 (not included; reuse the mlx-community MXFP4 repo).
FLUX.2 VAE: governed by its own (FLUX.2-dev) terms — not re-hosted here; the loader fetches it from source. Verify the VAE license for your use case.