File size: 3,027 Bytes
8e453f3 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 | ---
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).
|