File size: 1,815 Bytes
9d99afe | 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 | ---
license: apache-2.0
pipeline_tag: text-to-image
base_model: lodestones/Chroma1-HD
tags:
- mlx
- apple-silicon
- chroma
- flux
- text-to-image
---
# Chroma1-HD — MLX packed tiers (SceneWorks)
Pre-quantized, packed-load MLX tiers of [`lodestones/Chroma1-HD`](https://huggingface.co/lodestones/Chroma1-HD)
for on-device inference in [SceneWorks](https://github.com/SceneWorks) via `mlx-gen-chroma`.
Chroma1-HD is the high-detail full-CFG 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-HD`.
|