---
license: apache-2.0
base_model: Tongyi-MAI/Z-Image-Turbo
tags:
  - litert
  - tflite
  - text-to-image
  - on-device
  - diffusion-transformer
  - int8
pipeline_tag: text-to-image
library_name: litert
---

# Z-Image-Turbo — LiteRT bundle for Mirage iOS

Repackaged from [litert-community/Z-Image-Turbo-LiteRT](https://huggingface.co/litert-community/Z-Image-Turbo-LiteRT)
(Apache-2.0, converted from [Tongyi-MAI/Z-Image-Turbo](https://huggingface.co/Tongyi-MAI/Z-Image-Turbo))
with the host-side tensors required to run the full text-to-image pipeline on device:

| File | Role |
|---|---|
| `qwen_enc.tflite` | Qwen3-4B text encoder (penultimate hidden state) |
| `z_embx.tflite` / `z_refx.tflite` | image patch embed + noise refiner |
| `z_embc.tflite` / `z_refc.tflite` | caption embed + context refiner |
| `zc_main0..5.tflite` | 30 S3-DiT layers, 5 per chunk |
| `zc_final.tflite` | final adaLN + projection |
| `zvae.tflite` | VAE decoder |
| `host_tensors.safetensors` | `embed_tokens` (fp32), `t_embedder` MLP, `cap_pad_token`, `x_pad_token` |
| `tokenizer.json` | Qwen2 BPE tokenizer |

All graphs are INTEGER-int8, fixed 256×256 output, 64-token prompt budget.
The host loop (RoPE, adaLN timestep embedding, pad masking, x/cap concat,
flow-match Euler, VAE denorm `latents / 0.3611 + 0.1159`) runs in app code.

License: Apache-2.0 (inherited from Z-Image-Turbo).