Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,36 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
base_model: Tongyi-MAI/Z-Image-Turbo
|
| 4 |
+
tags:
|
| 5 |
+
- litert
|
| 6 |
+
- tflite
|
| 7 |
+
- text-to-image
|
| 8 |
+
- on-device
|
| 9 |
+
- diffusion-transformer
|
| 10 |
+
- int8
|
| 11 |
+
pipeline_tag: text-to-image
|
| 12 |
+
library_name: litert
|
| 13 |
+
---
|
| 14 |
+
|
| 15 |
+
# Z-Image-Turbo — LiteRT bundle for Mirage iOS
|
| 16 |
+
|
| 17 |
+
Repackaged from [litert-community/Z-Image-Turbo-LiteRT](https://huggingface.co/litert-community/Z-Image-Turbo-LiteRT)
|
| 18 |
+
(Apache-2.0, converted from [Tongyi-MAI/Z-Image-Turbo](https://huggingface.co/Tongyi-MAI/Z-Image-Turbo))
|
| 19 |
+
with the host-side tensors required to run the full text-to-image pipeline on device:
|
| 20 |
+
|
| 21 |
+
| File | Role |
|
| 22 |
+
|---|---|
|
| 23 |
+
| `qwen_enc.tflite` | Qwen3-4B text encoder (penultimate hidden state) |
|
| 24 |
+
| `z_embx.tflite` / `z_refx.tflite` | image patch embed + noise refiner |
|
| 25 |
+
| `z_embc.tflite` / `z_refc.tflite` | caption embed + context refiner |
|
| 26 |
+
| `zc_main0..5.tflite` | 30 S3-DiT layers, 5 per chunk |
|
| 27 |
+
| `zc_final.tflite` | final adaLN + projection |
|
| 28 |
+
| `zvae.tflite` | VAE decoder |
|
| 29 |
+
| `host_tensors.safetensors` | `embed_tokens` (fp32), `t_embedder` MLP, `cap_pad_token`, `x_pad_token` |
|
| 30 |
+
| `tokenizer.json` | Qwen2 BPE tokenizer |
|
| 31 |
+
|
| 32 |
+
All graphs are INTEGER-int8, fixed 256×256 output, 64-token prompt budget.
|
| 33 |
+
The host loop (RoPE, adaLN timestep embedding, pad masking, x/cap concat,
|
| 34 |
+
flow-match Euler, VAE denorm `latents / 0.3611 + 0.1159`) runs in app code.
|
| 35 |
+
|
| 36 |
+
License: Apache-2.0 (inherited from Z-Image-Turbo).
|