| --- |
| license: mit |
| language: |
| - en |
| pipeline_tag: text-to-image |
| library_name: mlx-gen |
| base_model: microsoft/Lens |
| tags: |
| - mlx |
| - apple-silicon |
| - lens |
| - text-to-image |
| - quantized |
| --- |
| |
| # Lens (base) — MLX pre-quantized tiers (SceneWorks) |
|
|
| Native-MLX, **pre-quantized** re-host of the **base** Lens model (`microsoft/Lens`, MIT) |
| for on-device Apple-Silicon inference via [`mlx-gen`](https://github.com/SceneWorks/mlx-gen)'s |
| `mlx-gen-lens` provider (SceneWorks). The heavy components are packed offline so a tier loads |
| directly with **no dense transient and no in-app quantization** (epic 8506, sc-8767). |
|
|
| Microsoft removed `microsoft/Lens` from the Hub; the base DiT here was recovered from the public |
| ungated re-package [`Comfy-Org/Lens`](https://huggingface.co/Comfy-Org/Lens) |
| (`diffusion_models/lens_bf16.safetensors`), whose keys are byte-identical to the diffusers |
| `LensTransformer2DModel` state dict. Base Lens and Lens-Turbo differ **only** in the DiT weights; |
| this re-host reuses the **shared** gpt-oss-20b text encoder + Flux.2 VAE + tokenizer + scheduler |
| from [`SceneWorks/lens-turbo-mlx`](https://huggingface.co/SceneWorks/lens-turbo-mlx). |
|
|
| Base Lens is **undistilled** — use a higher step count (~20–26) with CFG ~5.0 (the `mlx-gen-lens` |
| `lens` id defaults to 20 steps / CFG 5.0), unlike the distilled Turbo (4 steps / guidance 1.0). |
|
|
| ## Tiers |
|
|
| Each subdirectory is a full, self-contained turnkey snapshot (the diffusers multi-component tree — |
| `transformer/`, `text_encoder/`, `vae/`, `tokenizer/`, `scheduler/`, `model_index.json`): |
|
|
| | Tier | Dir | What is packed | |
| |------|-----|----------------| |
| | **Q4** (default) | `q4/` | DiT + gpt-oss encoder MoE experts → MLX group-64 affine 4-bit | |
| | **Q8** | `q8/` | DiT + gpt-oss encoder MoE experts → MLX group-64 affine 8-bit | |
| | **bf16** | `bf16/` | dense mirror of the source (no quantization) | |
|
|
| Two components are quantized (matching the load-time `.quantize` scope): |
|
|
| - **DiT** — `img_in`/`txt_in`/`proj_out` + every block's fused-QKV attention projections |
| (`img_qkv`/`txt_qkv`/`to_out.0`/`to_add_out`) and SwiGLU MLPs. The timestep embedder, AdaLN |
| modulations, and all norms stay full precision. |
| - **gpt-oss-20b encoder MoE experts** — the source ships these as MXFP4; the packed tiers store them |
| as MLX group-64 affine Q4/Q8 (stacked `experts.{gate_up,down}_proj.{weight,scales,biases}`). The |
| router / attention / embeddings / norms stay dense. |
|
|
| The **VAE** (the shared Flux.2 decoder) always runs f32 and is shipped dense in every tier. |
|
|
| The pack is **byte-identical** to what the load-time quantizer produces (bf16 cast, group 64), verified |
| in-repo (`mlx-gen-lens` `convert`/`quant` byte-identity tests) and by an on-device render gate. |
|
|
| ## License |
|
|
| MIT, inherited from `microsoft/Lens`. The shared text encoder is `openai/gpt-oss-20b` (Apache-2.0) |
| and the VAE is `black-forest-labs/FLUX.2-dev` (Apache-2.0). This is a format re-host; all model |
| weights and credit belong to the original authors (Microsoft Research; OpenAI; Black Forest Labs). |
|
|
