| --- |
| license: apache-2.0 |
| pipeline_tag: text-to-image |
| base_model: lodestones/Chroma1-Base |
| tags: |
| - mlx |
| - apple-silicon |
| - chroma |
| - flux |
| - text-to-image |
| --- |
| |
| # Chroma1-Base — MLX packed tiers (SceneWorks) |
|
|
| Pre-quantized, packed-load MLX tiers of [`lodestones/Chroma1-Base`](https://huggingface.co/lodestones/Chroma1-Base) |
| for on-device inference in [SceneWorks](https://github.com/SceneWorks) via `mlx-gen-chroma`. |
|
|
| Chroma1-Base is an 8.9B-parameter FLUX.1-schnell-derived text-to-image DiT (Apache-2.0). This repo |
| re-hosts it as three self-contained tiers so SceneWorks can load a tier directly with **no dense |
| transient and no in-app quantization** (the loader packed-detects group-wise affine weights via |
| `{base}.scales`): |
|
|
| | Tier subdir | Transformer weights | Notes | |
| |-------------|--------------------|-------| |
| | `bf16/` | dense bf16 | verbatim mirror of the source diffusers snapshot | |
| | `q8/` | packed Q8 (group 64) | transformer block Linears only | |
| | `q4/` | packed Q4 (group 64) | transformer block Linears only (default tier) | |
|
|
| **Quant scope.** Only the DiT `transformer/` matmul-heavy block Linears are quantized (the double |
| blocks' attention + FFN and the single blocks' attention + `proj_mlp`/`proj_out`). The transformer's |
| `x_embedder`/`context_embedder`/`proj_out` and the distilled-guidance Approximator, the shared T5-XXL |
| text encoder, and the FLUX.1 VAE stay dense in every tier. The Q4/Q8 packing is byte-identical to the |
| load-time quantization seam (weights cast to bf16 first, MLX group-wise affine at group size 64). |
|
|
| Each tier subdir is a complete diffusers-layout turnkey (`transformer/ text_encoder/ vae/ tokenizer/ |
| scheduler/ model_index.json`). |
|
|
| ## License |
|
|
| Apache-2.0, inherited from the upstream model. See `LICENSE`. Upstream: |
| `lodestones/Chroma1-Base`. |
|
|
