HiDream Sampler

One node, three clicks, a 17B model

When HiDream-I1 dropped in April 2025 it was the biggest open release in months - 17B parameters, MIT licence, GenEval scores above Flux.1-dev - followed immediately by the community's real problem: nobody could run it. Full precision wanted roughly 27GB of VRAM. What made it playable on a 16GB card was this pack, which loads the model in 4-bit NF4 and wraps the whole thing in a single node. That's HiDreamSampler: pick a model tier, type a prompt, hit generate. For a few months it was the standard way to run HiDream-I1 in ComfyUI, and the model comparisons people actually upvoted were generated with it.

The honest caveat comes from the author's own README: ComfyUI now supports HiDream natively, so on a current ComfyUI build you may not need this pack at all. It survives mostly because its sibling, HiDreamSamplerAdvanced, exposes encoder-level control you don't get from the native path, and because it auto-downloads and caches every weight. Think of this node as "HiDream-I1 the easy way," not "the only way."

How it works

Under the hood the node builds a full HiDream diffusion pipeline from a vendored fork of diffusers (hi_diffusers, shipped inside the pack), loads it 4-bit quantized, and runs sampling for you. The first generation downloads the weights from Hugging Face automatically - nothing to place by hand - then caches the pipeline in memory, so the second generation is dramatically faster. It swaps schedulers at generation time (UniPC for the full tier, a flow-matching Euler for dev/fast) and lets you switch to Karras variants if you want.

One thing to get straight early: HiDream-I1 is a 17B latent diffusion transformer. It is not HiDream-O1, the smaller pixel-space model the same lab shipped a year later. Anything you've read about O1 - no VAE, 8B, different hardware demands - does not transfer here.

The inputs that matter

Most of the required fields are self-explanatory, but a few are worth understanding before you click:

• model_type - full, dev, or fast, the three distilled tiers. Defaults to fast. full runs 50 steps with CFG 5.0; dev and fast are guidance-free (CFG 0) at 28 and 16 steps. Dev was the tier most of the community settled on.
• resolution - seven preset aspect ratios from 1024×1024 square up to 1248×832 landscape. override_width/override_height beat it if you set them to something other than 0.
• seed, num_images, prompt, negative_prompt - exactly what they sound like. num_images goes up to 8, generated in one batch from one seed.
• scheduler - Default for model is the right choice until you have a reason to leave it. UniPC, Euler, and the two Karras options are there for experimentation.
• override_steps / override_cfg / override_shift - leave at -1 and the node uses each model's tuned defaults. These exist so you can push past the built-in values; -1 literally means "use the model default."

The single output, image, is a standard IMAGE tensor - wire it into a PreviewImage or SaveImage node and you're done.

How to install

The easy route is ComfyUI Manager: search for "HiDream Sampler" and install. By hand:

cd ComfyUI/custom_nodes
git clone https://github.com/lum3on/comfyui_HiDream-Sampler
pip install -r comfyui_HiDream-Sampler/requirements.txt

Portable Windows users run the requirements line through their embedded Python instead. Then restart ComfyUI. The requirements list is heavy - transformers, diffusers, bitsandbytes, optimum, accelerate, gptqmodel, sageattention, and a pinned triton - and the weights (~15GB of NF4 builds, plus a 4-bit Llama text encoder) download automatically on first run. If you're in a region that can't reach Hugging Face, the README covers the HF_ENDPOINT=https://hf-mirror.com mirror workaround.

Troubleshooting

• Install is the pain point, historically. The pack's early versions depended on auto-gptq, which refused to build on Python 3.12; the fix that stuck was moving to gptqmodel. If you cloned an old copy, git pull and reinstall requirements so you get gptqmodel>=2.0.0.
• First generation looks stuck. It isn't - it's downloading the model. Watch the console for the Hugging Face progress.
• VRAM. NF4 builds fit in ~15GB; the full-precision tiers want ~27GB. Pick fast or dev on a 16GB card.
• The stop button doesn't work mid-generation - the author lists cancel support as unfinished. Don't queue long batches and wander off.
• It's a fading model. HiDream-I1 was effectively out of the conversation by early 2026 - no fine-tune ecosystem ever formed around a 17B base. If you want HiDream current, that's O1, and it's a different animal.