Iris-3B vs Qwen-Image 2.1 (Model Studio)

Iris-3B

Code and weights are Apache-2.0, and one Python package runs text-to-image, a 4-step Turbo, depth and 4x restoration from a documented CLI. It is ten days old, with no PyPI package, version tags, changelog or CI, needs an NVIDIA GPU and about 12 GB of weights, and has no server or API of its own.

Qwen-Image 2.1 (Model Studio)

qwen-image-2.1-pro generates and edits images with up to 10 references and transparent output, at $0.04 an image in Singapore, through an OpenAI-compatible endpoint or DashScope. Model Studio publishes a 99.9 per cent SLA and a no-training statement. The model allows 20 requests a minute, has no OpenAPI file, and its open weights are for non-commercial use only.

Iris-3B

1. Check for an NVIDIA GPU with CUDA and Python 3.11 or newer before downloading. The base weights are about 12 GB, and Qwen3-VL-4B downloads on the first run
2. Exclude depth/*, upscaler/* and turbo/* in hf download unless those models are needed. Each of the first two adds about 12 GB
3. Don't pass --steps, --cfg-scale, --negative-prompt, --order or --shift with a Turbo checkpoint. The script exits with an error
4. Pass --seed to reproduce an image, and set --outdir, since files are named after the prompt and a repeat run overwrites them
5. Load only exported directories (model.safetensors and config.yaml). A training .pth is read with torch.load(weights_only=False), so never load one from an untrusted source

Qwen-Image 2.1 (Model Studio)

1. Use the same region for the key, the model and the endpoint. Cross-region calls fail
2. Download each result within 24 hours. The URL expires and the image is purged
3. With the OpenAI SDK, pass image, seed, prompt_extend and watermark through extra_body, and write size as 1024x1024, not 1024*1024
4. Set a client timeout of up to 600 seconds for several outputs, or use the async DashScope call and poll the task id
5. Do not send prompt_extend_mode: agent with input images. It returns 400