Hunyuan 3 Generate (Telemetry)

"Hunyuan 3 Generate (Telemetry)" is the exact same generation node as the base Hunyuan 3 Generate - same inputs, same outputs, same speed - except it appends a RAM/VRAM usage report to the status output after every run. That's the entire difference. If you're debugging why a workflow OOMs or trying to figure out how much headroom a resolution actually needs, swap this node in and read the status string.

Given how VRAM-hungry HunyuanImage-3.0 is - an 80B MoE that wants roughly 45GB for NF4, 80GB+ for full BF16 - this is genuinely one of the more useful nodes in the pack. The pack already does aggressive memory management under the hood, but you can't tune what you can't see. This node makes the invisible visible.

What's different

Every input is inherited from the base generate node: model, prompt, seed, steps, resolution, guidance_scale, post_action, plus the optional LLM prompt-rewrite block (enable_prompt_rewrite, rewrite_style, api_url, model_name). Outputs are image, rewritten_prompt, status, and trigger.

The status output is the show here. After generation it carries the usual status text plus a memory snapshot - model size, VRAM used, and free RAM - formatted into the same string you can route into a Display String node or just read off the graph. Run the same workflow at a few resolutions and you'll quickly learn, on your hardware, what each preset costs.

When to reach for it

• You're chasing an OOM at 2MP+ and want to see where the peak hits.
• You're testing whether NF4 fits entirely on your card or is spilling to system RAM.
• You want to sanity-check that post_action: soft_unload_to_cpu actually freed the VRAM you expected.

Once you've dialed in your settings, there's no reason to keep it in the graph - the base Generate node runs identically. Treat Telemetry as a debugging instrument, not a permanent fixture.

Installing it

Same pack as everything else here, Comfy_HunyuanImage3:

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

Restart ComfyUI, then make sure a model is in ComfyUI/models/ (e.g. huggingface-cli download EricRollei/HunyuanImage-3-NF4-v2 --local-dir HunyuanImage-3-NF4). You'll find the Telemetry node right next to the base Generate node under the HunyuanImage3 category.

One honest warning

Telemetry gives you numbers, but don't treat them as gospel for a different GPU. The author's own benchmarks were run on a 96GB RTX 6000 Pro Blackwell, and the model's memory footprint scales with resolution (roughly ~15GB/MP of inference overhead at the high end). Use this node to characterize your own machine, not to predict someone else's. And if you're on a 24GB card, remember the fix is usually not more telemetry - it's the Low VRAM loader/generator pair instead.