Hunyuan MoE Efficient Test (A/B Quality)

The category says it all: this node lives under HunyuanImage3/Test. Hunyuan MoE Efficient Test (A/B Quality) (class HunyuanImage3MoETest) is the author's own quality-check harness for the memory-efficient MoE routing trick that powers the HighRes Efficient generate node - a way to generate with the standard and the efficient dispatch paths and eyeball whether they're actually identical. If you've ever wondered "sure, it's 75× less VRAM, but is it worse?", this is the node built to answer that question with your own eyes.

For most people this is not a daily-driver. You'll use it if you're evaluating the pack, comparing outputs, or just curious whether the loop-based expert routing sacrifices anything. The README's claim is that the efficient path makes identical routing decisions - same experts, same computations, same quality - and this node exists so you can verify that instead of taking it on faith.

How it works

It's a stripped-down generate node, deliberately. There's no offload_mode, no prompt rewrite, no resolution dropdown with 29 options - just the essentials:

• model (HUNYUAN_MODEL) - from any loader.
• prompt, seed, steps (default 40) - the usual.
• resolution - a fixed set of 9 near-native buckets (1024×1024, 1024×576, 576×1024, 1024×672, 672×1024, 1024×768, 768×1024, 1024×864, 864×1024). These are the aspect ratios the test is designed around.
• guidance_scale (default 6.0) - slightly lower than the generate nodes' 7.5 default, worth knowing if you're A/B-ing against a standard run and want apples-to-apples.
• post_action - keep_loaded (default), soft_unload_to_cpu, or full_unload, same as the other generators.

Outputs are image (IMAGE) and status (STRING).

How to install it

It ships with the pack:

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

Restart ComfyUI and hard-refresh the browser. Or search HunyuanImage in ComfyUI Manager.

Common issues & troubleshooting

The name says "A/B" but there's only one image output. Correct - you drive the A/B yourself by running the same prompt/seed through this node and through the standard Hunyuan 3 Generate, then comparing. Same seed, same steps, same resolution (pick one of the 9 buckets on both) and the only variable left is the MoE dispatch path. If the two images match to the eyeball, the efficient routing is doing what it claims.

This isn't the node for your real renders. It's a test harness with a locked-down resolution list and no offload controls. For actual high-res work, use Hunyuan 3 Generate (HighRes Efficient) - same efficient dispatch, full resolution range. Think of this node as the lab bench, and HighRes Efficient as the production tool that came out of the testing.

Guidance default difference. If your comparison looks different, double-check guidance_scale - 6.0 here vs 7.5 on the standard node will change output more than any dispatch difference. Set them equal for a fair test.