Can you fine-tune Qwen-Image-2.1-Turbo? Alibaba's pages disagree

Alibaba's qwen-image-2.1-turbo page lists capabilities as text and image input, image output and fine-tuning, but its general fine-tuning overview, read the same day, does not name the model among the ones it supports. Treat fine-tuning as unconfirmed until you see a Turbo option in your own Model Studio console, and test your look with reference images first.

Sume does not offer a fine-tuning route for images. Its image docs describe input_references instead, which shows a model your subject in the request itself.

What do the two Alibaba pages actually say?

The pages answer different questions, which is why they appear to conflict. The training overview's data thresholds are written for text models, with its SFT floor counted in question-answer entries, so do not assume the same numbers apply to an image model.

What should you decide before paying to train?

Fine-tuning an image model usually answers one of three needs: a consistent character or product, a house style, or a layout. Each can be tried first without training anything.

• Character or product: pass 1 to several clean photos as references and check whether the subject holds across ten generations.
• House style: pass two or three finished pieces as references and write the style rules in the prompt.
• Layout: use a fixed aspect_ratio or image_size and the same reference every time.

How do I run that test on Sume?

Sume's catalog lists Qwen Image, Qwen Image Max and Qwen Image Layered, and no Turbo row; see the closest-rows post. Read each model's own limits from the image models endpoint, because the docs say reference support is a per-model descriptor.

In the image generation docs, input_references is a range descriptor with a minimum and maximum per model. If a model's descriptor is {"min": 0, "max": 0}, it is text-to-image only and rejects references. Reference URLs must be public HTTPS; Sume rejects localhost, private-network and non-HTTPS URLs before submission.

Custom pixel sizes are accepted for Qwen through image_size as { width, height } or WIDTHxHEIGHT, per the same page. For a fair comparison, fix the prompt and the references, vary only the model, and score the outputs blind; the three versus ten reference test plan sets that up.

When is training still the right call?

If references cannot hold the subject, and you have rights to a few hundred images, training starts to make sense. Check the license first: the Hugging Face card names the Qwen Research License Agreement, and this post explains what that means for commercial output. Then confirm in your console that Turbo is selectable for training, and price the tuned model's per-image rate against the Turbo list price.