gufo/docs/models/qwen-image-2.1/README.md

Qwen-Image-2.1

Benchmarks · Quality · Experiments · ROCm migration

Native HIP image generation and editing on AMD Strix Halo (gfx1151). Loads BF16 safetensors directly; Python is not needed for inference. The checkpoint contains a diffusion transformer, Qwen3-VL prompt encoder and RGBA VAE.

Weights use the non-commercial Qwen Research License, separate from Gufo's MIT license. See the pinned sources and model contract.

Download about 33.1 GB and start the server:

MODEL_DIR=$(hf download Qwen/Qwen-Image-2.1 \
  --revision b3179ad355be050328e483a9dfdd9e60cd62adfa)
nix build
./result/bin/gufo serve image --model "$MODEL_DIR" \
  --served-model-name Qwen-Image-2.1 --port 8080

Generate an image:

curl --fail-with-body http://127.0.0.1:8080/v1/images/generations \
  -H 'Content-Type: application/json' \
  -d '{"model":"Qwen-Image-2.1","prompt":"A red ceramic teapot on a wooden table",
       "size":"1024x1024","seed":42}' > response.json

Edit with multiple references, in the order they are uploaded. For a single reference, keep just one image[] field:

curl --fail-with-body http://127.0.0.1:8080/v1/images/edits \
  -F model=Qwen-Image-2.1 \
  -F 'image[][email protected]' -F 'image[][email protected]' \
  -F 'prompt=Place the object from the first image in the room from the second image' \
  -F size=1024x1024 -F seed=42 > response.json

Both use the OpenAI Images response format. Save the first PNG:

jq -r '.data[0].b64_json' response.json | base64 --decode > output.png

Setting	Behavior
size	Default auto: 1024² generation; reference aspect ratio for edits. Explicit sides must be multiples of 32, at most 4096 per side and 4 megapixels total.
seed, steps	Seeded replay within the same build and request configuration. Default 40 steps; range 2–100. Fewer steps change quality.
n	1–10 outputs with independent seeds; requests execute serially on the GPU.
References	Up to 10 PNG/JPEG uploads, 20 MiB encoded and 32 megapixels combined; the HTTP body limit also applies.
background	auto, transparent, or opaque (composite over white).

Edit noise combines seed with the resized reference pixels. The same edit replays exactly, including after restarting the server, while editing an image with its generation seed no longer reuses its original noise. Seeded edits therefore differ from older Gufo releases; generation seeds are unchanged.

The official repository recommends 2048²; Gufo follows the pinned Diffusers 1024² default. Set size explicitly for larger output. Output is PNG/base64; URL output, separate masks, streaming and input_fidelity are unsupported. OpenAI SDK clients use the local /v1 base URL; seed and steps are Gufo extensions supplied through extra_body.

Qwen-Image-2.1-Turbo

The separately trained Turbo checkpoint (about 32.5 GB, same license) runs on the same engine, API and editing paths with its fixed eight-step schedule:

MODEL_DIR=$(hf download Qwen/Qwen-Image-2.1-Turbo \
  --revision fc13ba1986fcb905854dc8de2f52eee395f49b5b --exclude 'assets/*')
./result/bin/gufo serve image --model "$MODEL_DIR" \
  --served-model-name Qwen-Image-2.1-Turbo --port 8080

Omit steps or pass 8; other counts return HTTP 400. The checkpoint's eight sample_sigmas replace the base model's dynamic shift. It is about 4.8× faster than the 40-step base model at 1024²; see benchmarks and quality.

llama-swap

For llama-swap, add this model and send the same requests to the proxy:

models:
  Qwen-Image-2.1:
    cmd: >
      /path/to/gufo serve image --model /path/to/snapshot
      --served-model-name Qwen-Image-2.1 --port ${PORT}

For larger uploads, raise --max-request-bytes and the proxy's body limit. /health, /ready and /v1/models are available. Nix mkGufoServe supports modality = "image" with model and servedModelName.

Benchmarks · Quality · Experiments