# PromptEnhancer's memory table inverts at Q8_0

> Tencent Hunyuan's prompt rewriting tool for image generation, where the selection table lists the largest quantisation as needing more memory than the unquantised model, the low-memory downloads come from one contributor's personal account, the requirements file has no torch in it, and the default temperature makes the rewriting sampled rather than deterministic.

**Hunyuan-PromptEnhancer/PromptEnhancer** — [CVPR 2026] PromptEnhancer is a prompt-rewriting tool, refining prompts into clearer, structured versions for better image generation.

- Repository: https://github.com/Hunyuan-PromptEnhancer/PromptEnhancer
- Website: https://hunyuan-promptenhancer.github.io/
- Stars: 3,781 · Forks: 326
- Language: Python
- License: NOASSERTION
- Published: 2026-09-23 · Updated: 2026-09-23 · Language: en
- Canonical page: https://hysenlabs.com/projects/hunyuan-promptenhancer-promptenhancer

## The memory column inverts at the largest quantisation

The selection guide is a table with five rows and two numeric columns that do not agree with each other at the top. The 7B model is 13GB, rated high quality, and listed as needing 8GB or more of memory. The unquantised 32B model is 64GB, rated highest, and listed as 32GB or more. The largest GGUF quantisation of the same model, Q8_0, is 35GB, also rated highest, and listed as 35GB or more. Read those three together and the guide contradicts itself: on its own figures, choosing the highest quality quantisation costs more memory than choosing no quantisation at all. A tip below the table compounds it by claiming a 50 to 75 percent memory reduction from GGUF, which reads as a claim about download size rather than resident memory, and the table does not say so.

## The low-memory path runs through one contributor's account

Where the models live depends on which row of the table you are in. The full precision downloads come from the project's own organisation on Hugging Face, and the 7B comes from a subfolder of an entirely different repository belonging to the Hunyuan image model family. The four GGUF quantisations come from somewhere else again: a repository named after an individual contributor, under that person's account, with the quantisation named in the filename. The updates section credits the same person for adding GGUF support as a contribution. So the memory-efficient path that the page recommends for anyone without a data centre GPU depends on artefacts published by one outside contributor, while the full precision path depends on the vendor's own uploads.

## The model the page recommends starts with lives in another project's repository

For most users the recommendation is the 7B model, and the quick start downloads it from a `reprompt` subfolder inside the repository for a different Hunyuan image model, not from anything named after this project. Meanwhile the model the updates section tells you to try for higher quality enhancement has its badge commented out in the source of the readme, so a reader looking at the front page is shown neither the 7B nor the 32B as a link. Those links appear further down, in the download blocks, under two different organisations. So the front page shows a project, and the useful links are one screen down and split across three accounts.

## One model repository is spelled two ways in two adjacent references

The image-to-image model is named with an inconsistent capitalisation in two places a few lines apart. The badge row points at one spelling of the repository name, the download command below points at the other, and a third reference in the updates list uses the first spelling again. Repository names on the model hub are case sensitive, so one of those two forms will not resolve. The same section also links a hosted demo space under a third naming convention, with a low line between the words rather than a dash. None of this is fatal, and all of it is the kind of thing that costs an afternoon when the obvious fix is to copy the link.

## The requirements are unbounded and torch is not among them

The standard install is one line.

```bash
pip install -r requirements.txt
```

That file has seven entries and every one of them is a lower bound with no ceiling: a transformers floor, a datasets floor, a Google generative AI client floor, and four unbounded or version-light entries for the model hub, a vision-language utility package, requests and a progress bar. The notable absence is the deep learning runtime. The quickstart constructs the enhancer with an automatic device map, which is a transformers argument that exists to place tensors on a GPU, and the recommended quantised path is described as requiring CUDA support. Torch is not in the file, so it arrives from somewhere the page does not name.

## The default temperature means the rewriting is sampled

The text-to-image quickstart is four interesting lines long.

```python
    temperature=0.7,   # >0 enables sampling; 0 uses deterministic generation
    top_p=0.9,
    max_new_tokens=256,
```

Two things are documented in the comments rather than in the prose. First, the default temperature of 0.7 turns on sampling, and setting it to zero is what produces deterministic output, so the default behaviour of a prompt rewriter is to produce a different rewrite on different runs. Second, a comment above the call says the default system prompt is tailored for image prompt rewriting and can be overridden. The token budget of 256 is the other quiet constraint, and nothing on the page says what happens to a longer prompt.

## Two install paths, one of them a shell script

The second option is not a pip invocation.

```bash
chmod +x script/install_gguf.sh && ./script/install_gguf.sh
```

That path is described as being for quantised models with CUDA support, and a tip recommends it for faster inference with lower memory usage, especially for the large model. So a Python utility ships two installation mechanisms of different shapes, and the one aimed at constrained hardware is a bash script that the page never shows. The tree is small and says what the tool is: an `inference/` directory for the two entry point classes, a `models/` directory for downloaded weights, a `script/` directory for the installer, an `assets/` directory, and one evaluation script sitting at the root.

## A venue in the description and a preprint in the links

The repository description carries a conference tag in square brackets at the front, naming CVPR 2026, and the readme's own title gives the method as chain-of-thought prompt rewriting for text-to-image models. The only paper linked is an arXiv preprint identifier, and the updates list records the technical report as released in September 2025, which matches the identifier's year and month. No proceedings link appears anywhere on the page. The rest of the chronology is worth noting as a sequence: dataset and models first, both in September 2025, then the evaluation script in June 2026, nine months later. And the licensing is unresolved in the repository metadata, which records no recognised value, while a license file sits at the root.

## Conclusion

PromptEnhancer is worth looking at if you generate images and want the prompt expanded before it reaches the model, because the two modes and the model sizes are laid out clearly enough to choose from. Two things to check before trusting the selection guidance. The memory column contradicts the size column at the largest quantisation, so work it out from your own card rather than from the table. And the rewriting is sampled by default, with the code comment explaining that a temperature above zero enables sampling and that zero is what makes it deterministic, which means the same prompt will not come back rewritten the same way twice.

## FAQ

### what does prompt enhancer do

Hunyuan-PromptEnhancer restructures input prompts while preserving the original intent, producing clearer, structured prompts for downstream image generation. It supports both text-to-image generation and image-to-image editing, and for the editing case it refines the instruction with visual context from the input image.

### how to use prompt enhancer

Install with pip install -r requirements.txt or with the included GGUF shell script, download a model from the model hub, then construct the enhancer class with a models root path and a device map. Call its predict method with a prompt, a temperature, a top probability and a token limit; the enhancer accepts Chinese or English prompts.

### What is the best AI prompt enhancer?

The repository makes no comparison with other tools. It ships a 7B model recommended for most users and a 32B model for higher quality, plus four quantised variants of the 32B, in a table giving size, a qualitative quality rating, a memory floor and the hardware each suits.

### What is an example of a prompt?

The page's own quickstart uses a short English description, a third-person view of a race car speeding on a city track, and notes that the enhancer takes Chinese or English prompts. It does not discuss prompt examples more generally.

### vscode prompt enhancer

The repository says nothing about editor integration. It is a Python package with a pip install path, model downloads from a model hub, and two Python entry point classes, one for text-to-image prompt rewriting and one for image-to-image instruction refinement.

### how to use sulphur prompt enhancer

That is not a project in this repository. The page covers Hunyuan's PromptEnhancer, a prompt rewriting tool for text-to-image and image-to-image work, distributed as Python code plus downloadable model weights.

## Sources

- [Hunyuan-PromptEnhancer/PromptEnhancer on GitHub](https://github.com/Hunyuan-PromptEnhancer/PromptEnhancer)
- [Issues](https://github.com/Hunyuan-PromptEnhancer/PromptEnhancer/issues)
- [Project website](https://hunyuan-promptenhancer.github.io/)
- [README](https://github.com/Hunyuan-PromptEnhancer/PromptEnhancer/blob/main/README.md)

---

Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/hunyuan-promptenhancer-promptenhancer