Hunyuan 3 Generate (Low VRAM)

This is the generator you want when your Hunyuan 3 model doesn't fully fit on the GPU. "Hunyuan 3 Generate (Low VRAM)" is built for quantized NF4/INT8 models loaded with device_map offloading - the situation where the base Generate node breaks. It skips the CPU-offload calls that conflict with the loader's own offloading, and it hands control over to accelerate's device map instead of fighting it.

That distinction matters more than it sounds. The base Hunyuan 3 Generate assumes the model is fully resident on the GPU. Feed it a device_map-offloaded NF4 model on a 24GB card and you'll get the classic failure: "modules are dispatched on the cpu or the disk" - an error community members hit constantly when they first tried HunyuanImage-3.0 NF4 on 24–32GB cards. This node exists to not do that.

What it does differently

Inputs mirror the other generators - model, prompt, seed, steps, resolution, guidance_scale, offload_mode, post_action, plus the optional prompt-rewrite block - but the defaults tell the story:

• steps defaults to 40 (the tooltip's recommended range is 30–40).
• offload_mode defaults to smart, and the tooltip is refreshingly honest: for NF4/INT8 models this setting is ignored, because offloading is already handled by the loader's device map. Setting always on a quantized model just double-offloads and stalls.
• post_action defaults to soft_unload_to_cpu, which frees VRAM for downstream nodes after each run. On a memory-starved card, leaving the model parked on the GPU is a choice you rarely want to make.

Outputs are the usual four: image, rewritten_prompt, status, and trigger. Wire image to a Save/Preview node and you're done.

The resolution dropdown is worth a moment. Unlike the base node's ~1MP native buckets, this one includes consumer-friendly tiers up to 2.1MP (1080p landscape, 1440x1440 square) and a couple of 2K/3.7MP options. Stay at or under ~2MP on a 24GB card; past that you're asking a model that's already half-offloaded to also chew through a huge latent, and the waits get silly.

The pairing that matters

Low VRAM is only half the story - it expects a model from the NF4 Low VRAM+ loader (or an INT8 loader using device_map). The recommended 24–32GB setup is:

Hunyuan 3 Loader (NF4 Low VRAM+) → Hunyuan 3 Generate (Low VRAM) → Save Image

The loader's gpu_memory_target_gb slider keeps ~18–20GB free for inference, and this node works within that. Mixing a plain NF4 loader (which pins everything to GPU) with this node is legal but pointless - the base Generate node would be faster.

Installing

ComfyUI Manager → search "Comfy_HunyuanImage3", or:

cd ComfyUI/custom_nodes
git clone https://github.com/EricRollei/Comfy_HunyuanImage3
cd Comfy_HunyuanImage3
pip install -r requirements.txt

Restart ComfyUI. Dependencies include bitsandbytes>=0.48.2 and accelerate>=1.2.1, which are what actually power the offloading - if your install predates those versions, upgrade them.

Reality check

Low VRAM doesn't mean fast. Community testing shows NF4 on a 24GB 4090 works, but the offloaded weights stream over PCIe continuously and generation time balloons; one tester called it "not worth the time" on a 24GB card, while a 96GB card sailed through. The node removes the crashes, not the physics. If you can budget a 48GB+ card, use the plain Generate node and keep this one in your back pocket.