Added Many improvements!

• Added "use_uncensored_llm" option - this currently loads a different llama3.1-8b model that is just as censored as the first model. I will work on setting up a proper LLM replacement here, but may take a few days to get working properly. Until then this is just a "try a different LLM model" button. ** THIS IS STILL A WIP, DON'T @ ME **
• Renamed the existing node to "HiDream Sampler"
• Added new Node "HiDream Sampler (Advanced)"
• Exposed negative prompt for both (Note - only works on Full and Full-nf4 models)
• changed resolution to discrete integers to allow for free-range resolution setting instead of predefined values
• • Modified library to remove cap on output resolution/scale. Watch out, you can OOM yourself if you go too big. Also, I haven't tested really large images yet, the results will likely be really funky.
• modified the HiDream library to accept different max lengths per encoder model, as the previous 128 limit being enforced for all 4 encoder models was ridiculously low and stupid to put in front of an LLM or t5xxl.
• • For the 'simple' sampler node, I set the defaults to: CLIP-L: 77, OpenCLIP: 150, T5: 256, Llama: 256
• Advanced sampler node adds discrete values for max input lengths for each encoder, as well as a prompt box for each encoder. - - Default behavior is the primary prompt is used for all 4 encoder inputs unless overridden by an individual encoder input prompt.
• • • you can 'blank out' any encoder you don't want to use by simply leaving the primary prompt blank, then inserting a prompt for only the encoder(s) you want to use, or use different prompts for different encoders.
• • I think there is something wonky going on with the LLM encoder, it seems to have a lot of output 'noise' even when the input prompt is zeroed out, which I suspect is the LLM hallucinating output, will investigate but until then, the LLM encoder is way stronger than all of the other encoders, even when you don't feed it a prompt.

I will continue to add more improvements as I go, this has been fascinating to explore this model. I think there is still a lot of room for improvement (and optimizing).

Forked from original https://github.com/lum3on/comfyui_HiDream-Sampler

• Added NF4 (Full/Dev/Fast) download and load support
• Added better memory handling
• Added more informative CLI output for TQDM

Many thanks to the folks who created this and set it up for Comfy, I just spent a few hours adding better support for consumer GPUs.

• Full/Dev/Fast requires roughly 27GB VRAM
• NF4 requires roughly 15GB VRAM

HiDreamSampler for ComfyUI

A custom ComfyUI node for generating images using the HiDream AI model.

Features

• Supports full, dev, and fast model types.
• Configurable resolution and inference steps.
• Uses 4-bit quantization for lower memory usage.

Installation

Basic installation.

1. Clone this repository into your ComfyUI/custom_nodes/ directory:

   git clone https://github.com/lum3on/comfyui_HiDream-Sampler ComfyUI/custom_nodes/comfyui_HiDream-Sampler

2. Install requirements

   pip install -r requirements.txt

3. Restart ComfyUI.

Attention Installation

We've implemented four attention modules considering different hardware: FlashAttention 3, FlashAttention 2, SageAttention and Pytorch SDPA.

Attention Module	GPU
Flash-Attention 3	NVIDIA Hopper GPU
Flash-Attention 2	NVIDIA Ampere GPU and above
SageAttention 1	NVIDIA CUDA GPU
PyTorch SDPA	GPUs which support pytorch 2 (Including AMD GPU, Intel GPU)

We strongly recommend to install Flash Attention 2 with CUDA versions 12.4. Here's steps to install Flash-Attention 2:

• Get Flash-Attention 2 wheel from HuggingFace
• Install it in ComfyUI (.\python_embeded\python.exe -s -m pip install file.whl for portable version)
• Install accelerate .\python_embeded\python.exe -s -m pip install accelerate>=0.26.0
• Install this node with ComfyManager (or manually, don't forget to call python_embeded etc for portable version)
• add --use-flash-attention in "run_nvidia_gpu.bat"
• Use the "HiDream Sampler" node once to download the model

(If you don't want to install random wheel, you can take it from from here (it should create a release once it finish, which should take ~2 hours on GitHub CI))

Steps to install SageAttention 1:

• Install triton. Windows built wheel, download here:

.\python_embeded\python.exe -s -m pip install (Your downloaded whl package)

linux:

python3 -m pip install triton

• Install sageattention package

.\python_embeded\python.exe -s -m pip install sageattention==1.0.6

PyTorch SDPA is automantically installed when you install PyTorch 2 (ComfyUI Requirement). However, if your torch version is lower than 2. Use this command to update to the latest version.

• linux

python3 -m pip install torch torchvision torchaudio

• windows

.\python_embeded\python.exe -s -m pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu124

Download the weights

Here's some weight that you need to download (Which will be automantically downloaded when running workflow). Please use huggingface-cli to download.

• Llama Text Encoder

Model	Huggingface repo
4-bit Llama text encoder	hugging-quants/Meta-Llama-3.1-8B-Instruct-GPTQ-INT4
Uncensored 4-bit Llama text encoder	hugging-quants/Meta-Llama-3.1-8B-Instruct-GPTQ-INT4
Original Llama text encoder	nvidia/Llama-3.1-Nemotron-Nano-8B-v1

• Quantized Diffusion models (Thanks to azaneko for the quantized model!)

Model	Huggingface repo
4-bit HiDream Full	azaneko/HiDream-I1-Full-nf4
4-bit HiDream Dev	azaneko/HiDream-I1-Dev-nf4
4-bit HiDream Fast	azaneko/HiDream-I1-Fast-nf4

• Full weight diffusion model (optional, not recommend unless you have high VRAM)

Model	Huggingface repo
HiDream Full	HiDream-ai/HiDream-I1-Full
HiDream Dev	HiDream-ai/HiDream-I1-Dev
HiDream Fast	HiDream-ai/HiDream-I1-Fast

You can download these weights by this command.

huggingface-cli download (Huggingface repo)

For some region that cannot connect to huggingface. Use this command for mirror.

Windows CMD

set HF_ENDPOINT=https://hf-mirror.com

Windows Powershell

$env:HF_ENDPOINT = "https://hf-mirror.com"

Linux

export HF_ENDPOINT=https://hf-mirror.com

Usage

• Add the HiDreamSampler node to your workflow.
• Configure inputs: model_type: Choose full, dev, or fast. prompt: Enter your text prompt (e.g., "A photo of an astronaut riding a horse on the moon"). resolution: Select from available options (e.g., "1024 × 1024 (Square)"). seed: Set a random seed. override_steps and override_cfg: Optionally override default steps and guidance scale.
• Connect the output to a PreviewImage or SaveImage node.

Requirements

• ComfyUI
• GPU (for model inference)

Notes

Models are cached after the first load to improve performance and use 4-bit quantization models from https://github.com/hykilpikonna/HiDream-I1-nf4. Ensure you have sufficient VRAM (e.g., 16GB+ recommended for full mode).

Credits

Merged with SanDiegoDude/ComfyUI-HiDream-Sampler who implemented a cleaner version for my originial NF4 / fp8 support.

• Added NF4 (Full/Dev/Fast) download and load support

• Added better memory handling

• Added more informative CLI output for TQDM

• Full/Dev/Fast requires roughly 27GB VRAM

• NF4 requires roughly 15GB VRAM

Build upon the original [HiDream-I1]https://github.com/HiDream-ai/HiDream-I1