# Examples, measurements, and settings

## Downloads and presets

| Option | Generator | Complete text-to-image files | Runtime flags |
|---|---:|---:|---|
| Original reference (upstream) | 14.230 GB | 19.934 GB | `--diffusion-fa` |
| Q8 | 7.592 GB | 13.295 GB | `--diffusion-fa` |
| Q4 | 4.051 GB | 9.754 GB | `--diffusion-fa` |
| Q4 optimized | same Q4 file | 9.754 GB | `--fa --vae-conv-direct` |

Q4 generator weights are 71.5% smaller than the original. The complete Q4 text-to-image stack is 51.1% smaller than the original-generator stack with the same companions. Sizes are decimal GB and are not RAM requirements. Q4 optimized is a tested runtime preset, not a separate model or a new distillation. No calibrated weights, approximate VAE, or explicit cache quantization was used.

Download `Qwen-Image-2.1-Q4_K_S.gguf` or `Qwen-Image-2.1-Q8_0.gguf`, plus `Qwen3VL-8B-Instruct-Q4_K_M.gguf` (5.028 GB) and `qwen_image_2.1_vae_bf16.safetensors` (0.676 GB) from [`off-grid-ai/Qwen-Image-2.1-Companions-GGUF`](https://huggingface.co/off-grid-ai/Qwen-Image-2.1-Companions-GGUF). Original generator weights are linked in provenance, not duplicated here. The separate encoder retains Apache-2.0; it does not change the generator's research license.

## Matched image comparisons

1024×1024, seed 42, Euler, CPU RNG, VAE tiling. All variants share the same Q4 text encoder and original BF16 VAE. Thus the original control is not an entirely full-precision pipeline. Base uses 40 steps and CFG 6 with the engine default resolution-dependent schedule.

![Tabletop comparison](examples/tabletop-comparison.jpg)

![Portrait comparison](examples/portrait-comparison.jpg)

Full-resolution PNGs are in `examples/`; these comparison sheets are resized previews. Prompts and exact command arrays are included.

## Measurements

**Single runs on a shared GPU with other training jobs active. Elapsed times are observations under shared load, not controlled speed comparisons or estimates of idle-GPU performance. No speed ranking is established.** Times cover process launch, loading, inference, VAE decoding, and PNG output. GPU peaks were sampled once per second. These two synthetic prompts are examples, not a standardized benchmark suite. No mobile or laptop performance was measured.

| Scene | Variant | Time (s) | Peak GPU (GiB) | Host process high-water mark (GiB) |
|---|---|---:|---:|---:|
| tabletop | Q4 optimized | 115.110 | 9.77 | 1.54 |
| tabletop | Original | 96.797 | 19.31 | 2.32 |
| tabletop | Q8 | 78.670 | 13.07 | 1.61 |
| tabletop | Q4 | 80.455 | 9.77 | 1.51 |
| portrait | Original | 73.444 | 19.24 | 2.31 |
| portrait | Q4 optimized | 91.698 | 9.69 | 1.41 |

Hardware: RTX PRO 6000 Blackwell Server Edition. Engine binary SHA-256: `e6661c1908a586a1352a07f7c2ffe9516e7c6bf1927bd6c50e5bc63a4280be77`; commit unavailable. Numerical and visible differences are expected after quantization. No lossless-quality, SOTA-parity, or universal speed claim is made. These tests cover text-to-image only; editing and transparent output are not validated in this release.

## Run

Use a stable-diffusion.cpp build with Qwen Image 2.1 support. GGUF format alone does not ensure compatibility. This preset was tested on CUDA; other backends may need different runtime flags. See the [engine guide](https://github.com/leejet/stable-diffusion.cpp/blob/master/docs/qwen_image_2.1.md).

```sh
python run_example.py --scene tabletop --variant q4_optimized --engine /path/to/sd-cli
```

Choose `q8` or `q4` for the tabletop controls. `portrait` is available with `q4_optimized`. Full command arrays for the original controls are included for comparison with the upstream weights.