Lens (base)
Self-contained diffusers-layout snapshot of Microsoft's Lens text-to-image model, re-assembled for
in-house (SceneWorks) use after Microsoft removed the original microsoft/Lens repository from the Hub.
This is a repackage, not a retrain — every weight is byte-identical to a public upstream source (verified by tensor-level comparison against Comfy-Org's authentic redistribution):
| Component | Source | License |
|---|---|---|
transformer/ | Lens DiT, bf16 — from Comfy-Org/Lens (diffusion_models/lens_bf16.safetensors) | MIT |
text_encoder/ | gpt-oss-20b (MXFP4), used encoder-only — from openai/gpt-oss-20b | Apache-2.0 |
tokenizer/ | gpt-oss tokenizer — from openai/gpt-oss-20b | Apache-2.0 |
vae/ | FLUX.2 VAE (AutoencoderKLFlux2) — from black-forest-labs/FLUX.2-dev | FLUX.2-dev license |
The Lens text encoder is stock, frozen gpt-oss-20b (tensor-verified identical to Comfy's
gpt_oss_20b_nvfp4 up to quantization), and the VAE is stock FLUX.2-dev (full-file identical to
Comfy's flux2-vae). Only the transformer/ DiT is Lens-specific.
Layout
tokenizer/ tokenizer.json, tokenizer_config.json, special_tokens_map.json
text_encoder/ model-0000*-of-00002.safetensors (MXFP4), model.safetensors.index.json, config.json
transformer/ lens_bf16.safetensors
vae/ diffusion_pytorch_model.safetensors, config.json
Sampling defaults for the base model: 20 steps, guidance 5.0. For the distilled variant see
SceneWorks/Lens-Turbo (4 steps, guidance 1.0).