Card: the root export is gone from main (revision 216f2841 keeps it)
Browse files
README.md
CHANGED
|
@@ -23,9 +23,10 @@ This model has no row on [DeviceMark](https://devicemark.github.io/), the on-dev
|
|
| 23 |
<!-- gen-cards:devicemark end -->
|
| 24 |
|
| 25 |
> **2026-10-10 re-export, two folders.** `macos-fp16/` and `macos-int8/` were exported with apple/coreai-models
|
| 26 |
-
> 97a14be and load with its diffusion pipeline. The
|
| 27 |
-
>
|
| 28 |
-
>
|
|
|
|
| 29 |
|
| 30 |
# FLUX.2 klein 4B — Core AI
|
| 31 |
|
|
@@ -45,7 +46,7 @@ iPhone 17 Pro, over the ~6.1 GB a 12 GB iPhone gives one app.
|
|
| 45 |
| --- | --- | --- | --- | --- |
|
| 46 |
| `macos-fp16/` | fp16, one asset per graph | text-to-image | 19 | 14,159,882,995 |
|
| 47 |
| `macos-int8/` | int8 per-block 32 on the text encoder and the transformer, fp16 VAEs | text-to-image, image-to-image | 25 | 7,777,120,667 |
|
| 48 |
-
|
|
| 49 |
|
| 50 |
Text-to-image reads 16 files of a folder: 14,090,986,305 bytes from `macos-fp16/`, 7,540,032,209 bytes from
|
| 51 |
`macos-int8/`. The VAE encoders and the half-size VAEs are only for image-to-image and tiled decoding.
|
|
@@ -154,9 +155,9 @@ The text encoder is the Qwen3 model inside the FLUX.2 klein repo: hidden size 25
|
|
| 154 |
Qwen3-4B. Earlier versions of this card said 8B; that was wrong. The pipeline reads hidden states 9, 18 and 27, so the
|
| 155 |
export drops the last 9 layers.
|
| 156 |
|
| 157 |
-
## In-context editing (
|
| 158 |
|
| 159 |
-
|
| 160 |
**`Transformer_edit_2ref.aimodel`** for FLUX.2's in-context editing. You give a reference image and an instruction —
|
| 161 |
*"add a red wizard hat, keep everything else the same"* — and only the instructed change is applied while the
|
| 162 |
subject, pose, and background are preserved. This is different from strength-based image-to-image (SDEdit), which
|
|
|
|
| 23 |
<!-- gen-cards:devicemark end -->
|
| 24 |
|
| 25 |
> **2026-10-10 re-export, two folders.** `macos-fp16/` and `macos-int8/` were exported with apple/coreai-models
|
| 26 |
+
> 97a14be and load with its diffusion pipeline. The 2026-07-20 export that sat at the repo root (int4, plus the
|
| 27 |
+
> in-context edit transformers) was removed from `main` on 2026-10-10: current apple/coreai-models stops on it with
|
| 28 |
+
> "unsupported metadata_version", and no app reads it since the zoo's CoreAIImageGen app moved to the two folders.
|
| 29 |
+
> It is still there at revision `216f2841`.
|
| 30 |
|
| 31 |
# FLUX.2 klein 4B — Core AI
|
| 32 |
|
|
|
|
| 46 |
| --- | --- | --- | --- | --- |
|
| 47 |
| `macos-fp16/` | fp16, one asset per graph | text-to-image | 19 | 14,159,882,995 |
|
| 48 |
| `macos-int8/` | int8 per-block 32 on the text encoder and the transformer, fp16 VAEs | text-to-image, image-to-image | 25 | 7,777,120,667 |
|
| 49 |
+
| revision `216f2841` and earlier (2026-07-20; removed from `main` 2026-10-10) | int4 per-block 32, plus the in-context edit transformers | no app | 34 | 10,854,216,650 |
|
| 50 |
|
| 51 |
Text-to-image reads 16 files of a folder: 14,090,986,305 bytes from `macos-fp16/`, 7,540,032,209 bytes from
|
| 52 |
`macos-int8/`. The VAE encoders and the half-size VAEs are only for image-to-image and tiled decoding.
|
|
|
|
| 155 |
Qwen3-4B. Earlier versions of this card said 8B; that was wrong. The pipeline reads hidden states 9, 18 and 27, so the
|
| 156 |
export drops the last 9 layers.
|
| 157 |
|
| 158 |
+
## In-context editing (revision `216f2841`, 2026-07-20)
|
| 159 |
|
| 160 |
+
Revision `216f2841` and earlier have **`Transformer_edit.aimodel`**, **`Transformer_edit_512.aimodel`** and
|
| 161 |
**`Transformer_edit_2ref.aimodel`** for FLUX.2's in-context editing. You give a reference image and an instruction —
|
| 162 |
*"add a red wizard hat, keep everything else the same"* — and only the instructed change is applied while the
|
| 163 |
subject, pose, and background are preserved. This is different from strength-based image-to-image (SDEdit), which
|