Quick Verdict
PromptEnhancer is Tencent Hunyuan’s prompt-rewriting product family, not a page for one 7B or 32B model version. It covers text-to-image prompt enhancement and image-aware rewriting of image-editing instructions. The repository and weights are publicly downloadable, but they use the Tencent Hunyuan Community License rather than an OSI-approved open-source license. Because that agreement restricts territory, output use, very-large-MAU products, and AI training, this directory does not mark PromptEnhancer as recommended.
Best For
It fits research teams studying text-image alignment, developers standardizing prompts before generation, and organizations with suitable GPU capacity and license-review processes. A design team can use it as a preprocessing layer and compare the rewritten prompt across generators. It is not the easiest choice for a user seeking an instant browser image tool, nor for a product that cannot accept territorial and output restrictions. Large organizations must assess MAU across the licensee’s relevant products and services, not only this feature.
Key Features
- Rewrites short text-to-image prompts into more explicit descriptions of subjects, actions, style, layout, and attributes.
- Refines image-editing instructions using the source image as visual context.
- Tries to preserve intent while improving structure for downstream image models.
- Supports standard Transformers inference and GGUF quantized deployment paths.
- Includes research components such as Chain-of-Thought rewriting, AlignEvaluator, KeyPoints evaluation, and an evaluation script.
- Can run locally or behind a team-hosted API, with materially different privacy responsibilities.
Use Cases
Use PromptEnhancer to turn a vague creative brief into a more testable image prompt, clarify which elements an image edit should retain or replace, or run A/B evaluations of prompt adherence. It can also provide a visible rewrite stage inside an internal creative application. The rewrite still needs human approval: added details can shift composition, introduce unsupported claims, or narrow an intentionally ambiguous art direction.
Pricing
| Form | Software or weight price | Operational issue |
|---|---|---|
| Text-to-image family | Free to obtain | Hardware varies across full and quantized weights |
| Image-editing family | Free to obtain | Processes source images, increasing privacy sensitivity |
| Self-hosted API | Compute and operations | The deployer owns GPU, storage, security, and deletion controls |
| Third-party hosting | Operator-defined | Check price, logs, retention, training use, and data location separately |
The Tencent Hunyuan Community License applies only in its defined Territory, excluding the European Union, United Kingdom, and South Korea. If all products or services made available by or for a licensee exceeded 100 million monthly active users in the preceding month on the release date, a separate Tencent license is required. The agreement also prohibits using the works, outputs, or results to improve AI models other than Tencent Hunyuan or its derivatives.
Pros
- A focused rewriting layer that can precede different image generators.
- Covers text-to-image and image-editing workflows rather than one model release.
- Standard and GGUF routes offer quality-versus-memory choices.
- Published evaluator, keypoint framework, and scripts support systematic testing.
- Local inference can reduce dependence on unknown hosted prompt tools.
Cons
- The community license is not a standard open-source license and directly limits deployment territory and use.
- Outputs cannot be used to improve unrelated AI models, constraining synthetic-data and distillation workflows.
- Products above the specified 100M-MAU threshold require separate permission.
- Weights demand substantial downloads, VRAM, latency, and operations despite having no software purchase price.
- Hosted deployments need authentication, tenant isolation, log redaction, and deletion controls for prompts and images.
- Rewrites may drift from the brief, so prompts and generated images still require rights, factual, and brand review.
Alternatives
| Alternative | Better when | Tradeoff |
|---|---|---|
| ComfyUI | You need a visual graph for the entire generation pipeline | More workflow freedom, but prompt enhancement must be assembled separately |
| Stable Diffusion | You need a broad local image-generation ecosystem | A generation family, not a dedicated prompt-rewriting product |
| Midjourney | Direct image quality and aesthetic consistency matter most | Easier output flow, with different hosting and control boundaries |
| Ideogram | Typography inside generated images is central | Productized generation rather than a self-hosted rewriting research stack |
| Leonardo AI | You want a browser creative suite and asset workflow | More finished UX, less local deployment and license-level control |
FAQ
Is PromptEnhancer one model version?
No. It is a family spanning text-to-image, image-editing, and several full or quantized deployment forms.
Is it open source?
Public source availability is not the same as OSI open source. Its license is the Tencent Hunyuan Community License with territory, scale, and use restrictions.
Why is it not recommended here?
The technical work is useful, but a generic open-source recommendation would obscure restrictions on regions, outputs, large-scale products, and training other AI models.
Does local inference guarantee privacy?
It avoids sending inputs to an external model endpoint by default, but operators still must secure logs, caches, file permissions, and shared GPU systems. Hosted APIs require additional controls.
Can its output train another AI model?
The license prohibits using the works, output, or results to improve AI models other than Tencent Hunyuan or its derivatives. Data pipelines should enforce that boundary.
Bottom Line
PromptEnhancer is a technically substantive prompt-rewriting family with both text-to-image and image-editing paths. Adoption should start with license review, then hardware and privacy design, and finally human evaluation for rewrite drift. Teams that cannot comply with its territorial, output, MAU, and AI-training restrictions should choose a tool or service with terms aligned to their intended deployment.