model card
Browse files
README.md
ADDED
|
@@ -0,0 +1,57 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
base_model:
|
| 4 |
+
- Boogu/Boogu-Image-0.1-Base
|
| 5 |
+
- Boogu/Boogu-Image-0.1-Turbo
|
| 6 |
+
- Boogu/Boogu-Image-0.1-Edit
|
| 7 |
+
pipeline_tag: text-to-image
|
| 8 |
+
tags:
|
| 9 |
+
- mlx
|
| 10 |
+
- apple-silicon
|
| 11 |
+
- text-to-image
|
| 12 |
+
- image-editing
|
| 13 |
+
- boogu-image
|
| 14 |
+
library_name: mlx-gen
|
| 15 |
+
---
|
| 16 |
+
|
| 17 |
+
# Boogu-Image-0.1 — native MLX
|
| 18 |
+
|
| 19 |
+
Pre-converted, native **MLX** weights of [Boogu-Image-0.1](https://huggingface.co/Boogu) for Apple
|
| 20 |
+
Silicon, loaded by [SceneWorks](https://sceneworks.io)' `mlx-gen` Rust engine — no PyTorch / Python at
|
| 21 |
+
inference. Boogu is a Lumina-Image-2.0 / OmniGen2-lineage flow-matching image model: a mixed
|
| 22 |
+
single/double-stream DiT (~10.3B) + a Qwen3-VL-8B condition encoder + the FLUX.1 16-channel VAE.
|
| 23 |
+
|
| 24 |
+
## Variants
|
| 25 |
+
|
| 26 |
+
Each subfolder is a complete snapshot — `transformer/` (DiT) + `mllm/` (Qwen3-VL encoder + tokenizer)
|
| 27 |
+
+ `vae/` — loadable directly.
|
| 28 |
+
|
| 29 |
+
| Folder | Source | Task |
|
| 30 |
+
|---|---|---|
|
| 31 |
+
| `base/` | Boogu-Image-0.1-Base | text-to-image (true-CFG) |
|
| 32 |
+
| `turbo/` | Boogu-Image-0.1-Turbo | text-to-image (DMD few-step, no CFG) |
|
| 33 |
+
| `edit/` | Boogu-Image-0.1-Edit | reference image + instruction → edited image |
|
| 34 |
+
| `base-bf16/` `turbo-bf16/` `edit-bf16/` | (above) | same models, **full precision** |
|
| 35 |
+
|
| 36 |
+
The bare-named folders are the **default Q8** build; the `-bf16` folders are the unquantized originals
|
| 37 |
+
for maximum quality / experimentation.
|
| 38 |
+
|
| 39 |
+
## Quantization (Q8 folders)
|
| 40 |
+
|
| 41 |
+
Group-wise affine **Q8, group size 32** — the DiT hidden size (3360) is divisible by 32, not the usual
|
| 42 |
+
64. Quantized: the DiT Linears + the Qwen3-VL **text-tower** Linears. Kept full precision: the FLUX.1
|
| 43 |
+
VAE (decode-precision-sensitive), the Qwen3-VL **vision tower** (runs f32), and the token-embedding
|
| 44 |
+
table.
|
| 45 |
+
|
| 46 |
+
Footprint (measured @ 1024²): Q8 ~35.5 GB peak → recommended **64 GB** Mac. (Q4 would be ~27–30 GB /
|
| 47 |
+
48 GB; not shipped here.) bf16 runs require more.
|
| 48 |
+
|
| 49 |
+
## Usage
|
| 50 |
+
|
| 51 |
+
Loaded by `mlx-gen-boogu` via `BooguPipeline::from_snapshot("<variant folder>")` (text-to-image,
|
| 52 |
+
Turbo few-step, and single-reference image edit). Used by SceneWorks Image Studio.
|
| 53 |
+
|
| 54 |
+
## License
|
| 55 |
+
|
| 56 |
+
**Apache-2.0**, inherited from the upstream `Boogu/Boogu-Image-0.1-{Base,Turbo,Edit}` checkpoints.
|
| 57 |
+
These are re-quantized/repackaged copies of those weights for MLX; all credit to the Boogu authors.
|