Image 2.1 turbo for the Snapdragon 8 Elite neural unit
Built with Qwen. A phone conversion of Qwen-Image 2.1 with the Viggle 4-step turbo LoRA merged in, compiled as QNN context binaries for the Hexagon v79 neural unit (Snapdragon 8 Elite, e.g. iQOO 13). It draws 768x768 pictures from a prompt in 4 steps, and edits a 768x768 picture from an instruction.
Non-commercial use only (Qwen RESEARCH LICENSE AGREEMENT, see LICENSE and NOTICE).
What is in it
| Files | What they do |
|---|---|
r0.bin ... r3.bin, qwen_table_q8.raw, qwen_tokenizer.json |
the prompt reader (Qwen3-VL-8B's language model, 8-bit weights) |
vis.bin |
the reader's vision tower, for edits (8-bit weights) |
sp0.bin ... sp31.bin |
the drawing model, one block per file (4-bit weights in blocks of 32) |
s0.bin, s1.bin, qwen2b_table_q8.raw |
the Small prompt reader (Qwen3-VL-2B's language model, 8-bit weights) |
proj.bin |
the Small reader's translator: maps the 2B's features to the 8B's (trained on 35k prompts) |
enc.bin, dec.bin |
taeqi2_1 picture encoder and decoder |
manifest.json |
file list with sizes and SHA-256, used by the app's download |
Two downloads share the drawing model: Standard (the 8B reader, picture editing, 14.8 GB) and Small (the 2B reader + translator, prompts only, 8.2 GB; pictures very close to Standard's, the prompt read in ~2 s instead of ~7 s). On the phone a picture takes about 1 min 20 s and an edit about 3 min; the files are loaded one at a time, so it needs about 2.6 GB of memory while drawing.
Changes from the original models
The weights were converted to ONNX, quantized (4-bit blocks for the drawing model, 8-bit per channel with a Hadamard rotation for the reader) and compiled for the neural unit; the turbo LoRA is merged into the drawing model's weights. Pictures may differ slightly from the original models'.
The app, and building these files yourself
These files are made for the Anima app, which downloads them itself.
The scripts that build them from the original model weights, with a step-by-step guide, are in
build/ in this repo.