Qwen-Image negative_prompt 500 characters vs a Sume image request

Alibaba's Qwen-Image text-to-image API accepts a negative_prompt of up to 500 characters, but Sume's POST /v1/images request table does not list a negative prompt field. On Sume, put what you want excluded into the positive prompt as plain statements, and read the model's supported_parameters before sending anything extra.

The Qwen side of this comparison comes from Alibaba Cloud's Qwen-Image text-to-image API page, read on 2026-10-02. The Sume side comes from the Image API. Where the two disagree, the Sume page decides what your request can contain.

What does the Qwen-Image API give you that Sume does not list?

Alibaba's page lists a handful of request parameters for Qwen-Image. Three matter here: a negative_prompt capped at 500 characters, a prompt_extend switch for smart prompt rewriting that defaults to true, and a watermark switch that adds Qwen-Image branding and defaults to false. The page also says the prompt itself is limited to 800 tokens, or 1,300 for the 2.0 series.

Sume's request table for images lists model, prompt, n, resolution, aspect_ratio, size, quality, output_format, mask_url, background, output_compression, seed, stream, input_references, the provider object, metadata, mode, webhook_url and wait_timeout_seconds. There is no negative prompt, no prompt-rewrite switch and no watermark switch in that list.

How do you exclude things on Sume without a negative prompt?

Describe the scene you want and state the exclusions as part of it. Instead of negative_prompt: "text, logo, extra fingers", write a prompt such as: "A clean product photo of a ceramic mug on a plain white table, no text, no logo, no people in frame." This is a prompt-writing technique, not a Sume feature, and it works with whatever model reads the sentence. It is not guaranteed the way a dedicated field would be, so check the output and regenerate when a stray element appears.

If the exclusion is the whole job, such as removing a watermark-like element from an existing picture, use a reference image and an edit-style prompt instead. Sume takes input_references as public HTTPS image URLs, and models whose descriptor is {"min": 0, "max": 0} are text-to-image only and reject references.

What happens if you send the field anyway?

The Image API page states that a request which sets a parameter the selected model does not list is rejected with 400 unsupported_parameter rather than silently dropped. For the named fields that no model advertises yet, seed and output_compression, the page says the same status applies. The page does not spell out the outcome for a key it has never defined, so treat an unknown key like negative_prompt as something to leave out, and test it once against a cheap model if you want to see the error shape for yourself.

Failed generations are not billed, per the page's billing section, so an experiment that comes back as a 400 costs nothing.

How do you check what a Qwen model on Sume accepts?

List the catalog and read the descriptors rather than copying a vendor's parameter list. The call below returns every model with its supported_parameters; look for qwen/qwen-image and qwen/qwen-image-max.

curl "https://api.sume.com/v1/images/models/qwen/qwen-image-max/endpoints" \
  -H "Authorization: Bearer $SUME_API_KEY"

What to do next

Pick the Qwen model on the catalog, put exclusions in the prompt, and keep n inside the range the descriptor prints. For the text-only behaviour of the Max model see Qwen Image Max on Sume, and for the question of newer Qwen releases see Qwen-Image 3.0 on Sume. Sume does not mirror Alibaba's prompt rewriting or watermark options, so if you depend on those, call Alibaba directly for that part of the workflow.

Qwen exclusions on Sume: a checklist

Prompt wording is a craft, not a contract. If a requirement is legal or brand-critical, such as never showing a logo, add a human or automated check after generation instead of trusting any prompt or negative field.

• Write what the picture should contain first, then add a short clause listing what must be absent, in the same prompt.
• Keep each exclusion concrete: say no visible text rather than no bad things, because a vague phrase gives the model nothing to act on.
• Regenerate rather than argue with the prompt: a failed generation is not billed, but a completed image you reject still is, so keep batches small while you tune.
• Log the full prompt in metadata on the job. The Image API page says metadata is stored on the job and not sent to the provider, which makes it a good place to record which exclusion wording produced which result.