What is Prompt rewriting (prompt extend)? Definition, key points and related tools
Automatic LLM expansion of your prompt (on by default on Wan, Qwen-Image and others). Turn it off for drama storyboards: the rewriter overrides cut points, durations and facing rules.
What the rewriter does, and why it collides with a storyboard
A prompt rewriter is a language model placed between your text and the image or video model. It exists because most prompts typed into a playground are one line long. Alibaba's Wan 3.0 API reference says the rewrite significantly improves quality for shorter prompts but increases latency. Its job is to add what a short prompt leaves out.
The additions are concrete. Alibaba's Qwen-Image reference shows a one-sentence cat prompt returned as a photorealistic wildlife portrait with a tilted head, perked ears and soft natural lighting. Microsoft's documentation for DALL-E 3 shows a watercolour of the Seattle skyline gaining a setting sun, the Space Needle and Mount Rainier. Pose, framing, time of day and light were all chosen by the rewriter, not the author.
A drama storyboard is the opposite of a short prompt: it already fixes shot boundaries and seconds, who is in frame, facing, quoted dialogue and light sources (see storyboard script). Every detail a rewriter adds either duplicates or overrides one you wrote on purpose, and it aims at one attractive clip, not at matching the previous segment. Rewriting is also non-deterministic (Microsoft states this for DALL-E 3), so one storyboard can be paraphrased differently on every call, bringing back the variance the storyboard was written to remove.
Where the switch lives, model by model
Official documentation, checked 2026-09-12:
| Model / platform | Setting | Default | Can you read the rewritten text? |
|---|---|---|---|
| Wan 3.0 (Alibaba Cloud Model Studio) | prompt_extend | true | No field documented; the response returns orig_prompt, your original text |
| Wan 2.x text-to-video | prompt_extend | true | wan2.5 and earlier return actual_prompt; wan2.6 never does, and the wan2.7 response documents orig_prompt alone |
| Qwen-Image | prompt_extend | true | Asynchronous responses include actual_prompt; the synchronous response documents no such field |
| MiniMax Hailuo 2.3 / 02 image-to-video | prompt_optimizer | true | Not documented; MiniMax says to set it to false for more precise control |
| Seedream (BytePlus ModelArk) | optimize_prompt_options.mode | standard; a faster mode exists on some versions, and no off value is listed | Not documented |
| Vidu start-end-to-video | is_rec | optional; default not stated | When true, a system-generated prompt replaces yours, for 10 extra credits per task |
| DALL-E 3 on Azure OpenAI | always applied | cannot be disabled | Yes, revised_prompt (the model was retired on Azure on 2026-03-04) |
Two more patterns matter. BytePlus documents that camera_fixed, listed for Seedance 1.0 pro and 1.5 pro, appends a fixed-camera instruction to your prompt with no guaranteed result, so your text changes without anything called a rewriter. The MiniMax H3 create-video reference documents no optimiser parameter at all.
When to leave it on, and how to check what it did
- Ideation and one-liners: leave it on. That is the input it was built for.
- Storyboard-driven production: set the flag to false explicitly in every request. Qwen-Image's own reference recommends false for tighter control over image details.
- Do not rely on defaults. They are set per model version, and the fields that come back change between versions (actual_prompt on wan2.5 and earlier, nothing on wan2.6, orig_prompt alone on wan2.7 and in the Wan 3.0 reference).
- Audit where you can. If the rewritten text is returned, diff it against the storyboard: count shot markers and timestamps, quoted lines, named characters and seconds. If it is not returned, run one segment twice with identical inputs, flag on and off, and compare cuts executed, lines spoken verbatim, extra people and light direction. Alibaba notes that even a fixed seed may not reproduce a Wan 3.0 result exactly, so judge on those concrete checks rather than on overall look.
- Want enrichment? Make it a visible step. Google's Veo 3.1 prompting guide suggests asking Gemini to enrich a simple prompt; you review the enriched text before it reaches the video model.
Common mistakes and their causes
- Judging the setting in a playground. Test prompts are short, so rewriting looks like a quality switch; production prompts are long and structured, where the same switch adds noise.
- Turning it off and keeping thin prompts. With rewriting off, nothing fills empty slots. The storyboard has to carry subject, scene, motion, camera, light and sound itself (see writing prompts for video models).
- Assuming off means verbatim. Options such as camera_fixed on Seedance 1.x still append text, Vidu's is_rec replaces the prompt, and safety screening runs regardless (see content moderation).
- Ignoring length limits. Wan 3.0 truncates prompts beyond 20,000 characters and MiniMax H3 caps each text item at 7,000; BytePlus recommends no more than 600 English words for Seedream, noting that long prompts make the model overlook details.
Key figures and sources
| Figure | Value | Source | Date |
|---|---|---|---|
| Wan 3.0 prompt_extend | default true; response returns orig_prompt (original text), no rewritten-prompt field | www.alibabacloud.com | 2026-09-12 |
| Qwen-Image prompt_extend | default true; actual_prompt documented for asynchronous responses, not synchronous; negative prompt not affected | www.alibabacloud.com | 2026-09-12 |
| MiniMax Hailuo 2.3 / 02 prompt_optimizer | default true; set false for more precise control | platform.minimax.io | 2026-09-12 |
| Vidu is_rec (recommended prompt) | ignores the entered prompt; 10 extra credits per task | platform.vidu.com | 2026-09-12 |
| DALL-E 3 prompt transformation (Azure OpenAI) | cannot be disabled; model retired 2026-03-04 | learn.microsoft.com | 2026-09-12 |
FAQ
Does turning prompt rewriting off lower quality?
For short prompts it often does; that is the case the feature targets, and Alibaba describes the gain as largest for shorter prompts. A full storyboard already fills the slots a rewriter would invent, so switching it off keeps your text as written. If output gets worse with it off, the storyboard is missing information: add it rather than letting a rewriter guess.
How can I see what the rewriter changed?
Where it is exposed: actual_prompt on wan2.5 and earlier and on asynchronous Qwen-Image calls, revised_prompt on DALL-E 3. The wan2.7 and Wan 3.0 references document orig_prompt alone, so compare two runs with the flag on and off. The MiniMax and BytePlus references checked on 2026-09-12 document no field for the optimised prompt.
Is prompt rewriting the same as a negative prompt or a safety filter?
No. The rewriter edits your positive prompt; Qwen-Image's reference states it does not touch the negative prompt, which is a separate field. Safety screening runs whether or not rewriting is on, so switching it off does not change what gets blocked (see content moderation).
Sources
- https://www.alibabacloud.com/help/en/model-studio/wan3-video-generation-api-reference
- https://www.alibabacloud.com/help/en/model-studio/text-to-video-api-reference
- https://www.alibabacloud.com/help/en/model-studio/qwen-image-api
- https://platform.minimax.io/docs/api-reference/video-generation-i2v
- https://platform.minimax.io/docs/api-reference/video-generation-v2-create
- https://docs.byteplus.com/en/docs/ModelArk/1541523
- https://docs.byteplus.com/en/docs/ModelArk/1520757
- https://platform.vidu.com/docs/start-end-to-video
- https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/prompt-transformation
- https://cloud.google.com/blog/products/ai-machine-learning/ultimate-prompting-guide-for-veo-3-1