# HunyuanDiT: a Chinese-prompt diffusion transformer whose README news stops in December 2024 and whose pins span two eras > Tencent-Hunyuan/HunyuanDiT holds the PyTorch model definitions, pretrained weights and sampling code for Hunyuan-DiT, a text-to-image diffusion transformer built around Chinese prompts. This article covers the versions, the speed-ups, the control add-ons and the dependency pins. **Tencent-Hunyuan/HunyuanDiT** — Hunyuan-DiT : A Powerful Multi-Resolution Diffusion Transformer with Fine-Grained Chinese Understanding - Repository: https://github.com/Tencent-Hunyuan/HunyuanDiT - Website: https://dit.hunyuan.tencent.com/ - Stars: 4,290 · Forks: 366 - Language: Jupyter Notebook - License: NOASSERTION - Published: 2026-10-16 · Updated: 2026-11-04 · Language: en - Canonical page: https://hysenlabs.com/projects/tencent-hunyuan-hunyuandit ## Prompts can be Chinese sentences, and a second turn can edit the first picture The title names the selling point: fine-grained Chinese understanding. The README's example prompts include 画一只穿着西装的猪, which it pairs with the English line draw a pig in a suit, and 生成一幅画，赛博朋克风，跑车, paired with generate a painting, cyberpunk style, sports car. Plain natural-language sentences are the intended input and no keyword syntax is described. Multi-turn use is the second feature. The example is 画一个木制的鸟, draw a wooden bird, followed by 变成玻璃的, turn into glass. The second message refers to the first image instead of restating it. A companion paper, DialogGen, covers multi-modal interactive dialogue for multi-turn text-to-image generation, and the May 22, 2024 news entry says the repository supports running the multi-turn demo through a Gradio script. The README also points to a hosted route, the Tencent Hunyuan Bot on the web. Prompts containing drawing-related keywords activate the text-to-image feature there, and the README says it is free. The repository contains PyTorch model definitions, pretrained weights and inference and sampling code for the paper, so the web bot and the local code are two separate ways in. The sampling entry points at the top level are `sample_t2i.py`, `sample_controlnet.py` and `sample_ipadapter.py`. ## Distillation and TensorRT speed-ups are claimed as 50 percent and 47 percent on NVIDIA GPUs Two acceleration paths come from the news list. On Jun 06, 2024 the project introduced a Distillation version of Hunyuan-DiT, which it says achieves 50% acceleration on NVIDIA GPUs. On May 22, 2024 it introduced a TensorRT version, with 47% acceleration on NVIDIA GPUs and instructions in a TensorRT-libs repository on Hugging Face. Both figures come without a stated baseline, GPU model or resolution in the README news entries, so they are not comparable to each other and not a benchmark you can reproduce from this page. The repository has a `trt/` folder, which is the code side of the TensorRT path. Memory is addressed separately. On Jun 27, 2024 the project released 6GB GPU VRAM inference scripts, found in the `lite/` folder. That is the route for a small consumer card. On older hardware, the Sep 13, 2024 note says scaled attention is used in place of flash attention on V100 GPUs. A V100 user therefore runs scaled attention rather than flash attention. The Distillation weights are published as Distillation, Distillation-v1.1 and Distillation-v1.2 on Hugging Face, matched to the base releases. ## Version 1.1 fixed oversaturation and watermarks, and 1.2 is the base for most add-ons Three base versions appear in the news. The original release is the baseline. HYDiT-v1.1, released Jun 13, 2024, mitigates the issue of image oversaturation and alleviates the watermark issue. HYDiT-v1.2 followed on Jul 08, 2024, with its own weights and its own distilled variant on Hugging Face. The version matters for everything built on top. The community ControlNet models in the contribution list are for HunyuanDIT v1.2: inpaint, tile and lineart. The ComfyUI update on Jul 15, 2024 supports weights for versions 1.1 and 1.2, including those trained by Kohya or the official script. An anime base model from a community member is also v1.2. If you pick the original weights, community add-ons built for 1.2 may not apply. Match the add-on to the version it names. Training is covered by code. The Jun 13, 2024 entry released full-parameter training and LoRA training. The Dec 17, 2024 entry optimizes LoRA training with a refined grad checkpoint and a low-bit optimizer, enabled by passing `--lowbit-opt`. Hunyuan-Captioner, released Jun 27, 2024 in the `mllm/` folder, produces fine-grained captions for training data. ## ControlNet covers canny, pose and depth, and IPAdapter was added in September 2024 ControlNet was released on Jun 19, 2024, supporting canny, pose and depth control, with training and inference code in the `controlnet/` folder. The community adds inpaint, tile and lineart controlnets for v1.2, hosted outside this repository on Hugging Face. IPAdapter followed on Sep 13, 2024, with its documentation in the `ipadapter/` folder. LoRA code is in `lora/`. The integration list is wide. Hunyuan-DiT is available in Hugging Face Diffusers since Jun 05, 2024, with LoRA and ControlNet support in diffusers added Jun 27, 2024, under the `diffusers/` folder. ComfyUI support arrived Jun 06, 2024 and was extended to ControlNet and LoRA on Aug 26, 2024. A Kohya version for training, with a GUI for inference, is in `kohya_ss-hydit/`, and the Jul 03, 2024 note says the official Kohya version is under review. Each integration lives in its own folder with its own documentation. A user who wants ControlNet in ComfyUI reads the `comfyui/` folder, not the main README. That structure keeps the base model small to describe, and it also scatters the setup steps across folders. ## requirements.txt pins 2023-era libraries next to torch 2.7.1 or newer The repository file requirements.txt mixes two generations of packages. It asks for `torch>=2.7.1` and `torchvision>=0.21.0`, both open-ended lower bounds. In the same file, `transformers==4.39.1`, `diffusers==0.21.2`, `timm==0.9.5`, `peft==0.10.0`, `accelerate==0.29.3`, `deepspeed==0.6.3`, `gradio==3.50.2` and `protobuf==3.19.0` are exact pins. The TensorRT-related lines are old too: `cuda-python==11.7.1`, `onnxruntime==1.12.1`, `onnx==1.12.0`, `onnx-graphsurgeon==0.3.27`, `polygraphy==0.47.1` and `nvidia-pyindex==1.0.9`, with an extra index at pypi.ngc.nvidia.com. Also pinned are `pyarrow==16.1.0`, `pandas==2.0.3`, `einops==0.7.0` and `sentencepiece==0.1.99`. The file does not say whether this combination has been installed together. Exact pins on a diffusion stack of that age, combined with a very recent torch floor, are where resolver conflicts show up, so expect to adjust it for your GPU. The repository also ships `environment.yml`, and the Jul 15, 2024 news entry offers Docker environments for CUDA 11 and CUDA 12, meant to bypass complex installations. If the pip route fights you, the Docker route is the one the project itself suggests. ## The README has no install or run commands, and GitHub reports the license as NOASSERTION Installation, pretrained-weight download and inference instructions are in later README sections, which the news entries link to through anchors such as the installation guide for Linux, the download section and the Gradio section. The repository layout shows what exists. Top-level folders include `hydit/` for the model, `IndexKits/` and `dataset/` for data handling, `app/` and `utils/`, `tests/`, and `asset/`. A `.gitmodules` file is present. The license field reads NOASSERTION. The repository has a LICENSE.txt and a Notice file. A team planning commercial use should read both files rather than assume a standard open-source license. The primary language shows as Jupyter Notebook, which reflects notebook files in the repository and not the core code. The model code itself is PyTorch, in `hydit/`. ## The newest README news item is from December 2024, though the last push was 2026-10-08 The repository is not archived, and its last push was on 2026-10-08. It has no GitHub releases. The most recent news line in the README is dated Dec 17, 2024, about the LoRA optimization. Nothing in the news list reports a change after that date, so the README itself reads as a 2024 snapshot. Some entries have also aged out. The Jul 15, 2024 fine-tuning event with Shakker.Ai, which offered up to $230 in bonuses, is listed in the news with no end date given. Do not read it as a current offer. The Community Contribution Leaderboard credits outside work: TTPlanetPig for the v1.2 inpaint, tile and lineart ControlNets and ComfyUI nodes, sdbds for Kohya training tools, CrazyBoyM for ComfyUI ControlNet support, and L_A_X for an anime base model. These live in other repositories and Hugging Face pages, so their upkeep is separate. The Open-source Plan checklist marks Inference, Checkpoints and a Distillation version as done. Treat HunyuanDiT as a finished 2024 release line with a few later touches, and check each add-on's own page before depending on it. ## Conclusion Choose HunyuanDiT if you need a text-to-image model that takes Chinese prompts and want ControlNet, LoRA and ComfyUI support from the same project. Avoid it if you need clear license terms, because GitHub reports no recognised license and the LICENSE.txt and Notice files must be read directly. Before installing, check requirements.txt against your CUDA and PyTorch setup. ## FAQ ### What is HunyuanDiT? Hunyuan-DiT is a multi-resolution diffusion transformer for text-to-image generation with fine-grained Chinese understanding. The repository has the PyTorch model definitions, pretrained weights and inference and sampling code. ### Which HunyuanDiT versions exist? The news lists the original release, v1.1 from Jun 13, 2024 and v1.2 from Jul 08, 2024. Version 1.1 mitigates oversaturation and the watermark issue, and each version has a distilled variant on Hugging Face. ### Can HunyuanDiT run on a small GPU? The project released 6GB GPU VRAM inference scripts on Jun 27, 2024, in the lite folder. It also lists Distillation and TensorRT versions for NVIDIA GPUs. ## Sources - [Issues](https://github.com/Tencent-Hunyuan/HunyuanDiT/issues) - [Project website](https://dit.hunyuan.tencent.com/) - [README](https://github.com/Tencent-Hunyuan/HunyuanDiT/blob/main/README.md) - [Tencent-Hunyuan/HunyuanDiT on GitHub](https://github.com/Tencent-Hunyuan/HunyuanDiT) --- Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/tencent-hunyuan-hunyuandit