#5858th of 90 in Open modelsFor image, video and design, image and design
promptenhancer
Prompt-rewriting LLMs that expand short prompts for text-to-image and image editing
Live demo ↗
(opens in a new tab)Website ↗
(opens in a new tab)Repository on GitHub ↗
(opens in a new tab)
- Stars
- 3.8k
- License
- custom license
- Last commit
- Aug 2026
- Language
- Python
3,659 on 25 Apr 20263,808 on 14 Nov 2026
From GitHub's daily star counts.
Checked on GitHub.
Overview
Tencent Hunyuan's PromptEnhancer rewrites short or vague image prompts into structured, detailed ones while keeping the original intent, using chain-of-thought rewriting. It ships 7B and 32B Transformers models, an Img2Img editing variant built on Qwen2.5-VL that reads the input image, and quantized 32B GGUF files that run through llama.cpp. Python classes expose a predict() call for each mode.
Who it is for: Developers adding prompt rewriting in front of image-generation models
Strengths
- Covers both text-to-image prompts and image-editing instructions that use the input image as context
- Offers 7B, 32B and GGUF Q8_0, Q6_K and Q4_K_M, with a stated memory figure for each
- Includes the T2I-Keypoints-Eval dataset and evaluation script, plus a technical report on arXiv
- Small Python API: one predict() call with temperature, top_p and max_new_tokens
Weaknesses
- License is not detected by GitHub (NOASSERTION), so reuse terms need checking before adoption
- 32B models need 20-64GB of memory; only the 7B fits typical consumer GPUs
- Library only: no HTTP server, UI or Docker setup, so serving must be built separately
- GGUF path handles text prompts only; image editing needs the 32B Transformers model
What it needs
- RAM ≥ 8 GB
- GPU optional
- Needs PyTorch/Transformers, llama.cpp (GGUF), Hugging Face CLI
- Models: PromptEnhancer-7B, PromptEnhancer-32B, PromptEnhancer-Img2Img-Edit (Qwen2.5-VL), GGUF Q8_0/Q6_K/Q4_K_M
Also in Open models
See all 90| Rank | Project | Score |
|---|---|---|
| 1 | ruviewWiFi CSI sensing on ESP32 for presence, breathing, heart rate and pose | 94 out of 100 |
| 2 | layaSingle-pass multilingual classifier returning typed decisions with calibrated odds | 88 out of 100 |
| 3 | rf-detrReal-time transformer models for object detection, segmentation and keypoints | 82 out of 100 |
| #4 | timesfmPretrained time-series foundation model for zero-shot forecasting from Google Research | 80 out of 100 |
| #5 | open_clipPyTorch library for training and running CLIP-style image-text models | 79 out of 100 |