Hunyuan Latent Noise Shaping

Hunyuan Latent Noise Shaping (class HunyuanLatentNoise) is part of the pack's experimental latent-control family - the nodes that try to give the base (non-Instruct) Hunyuan pipeline the kind of composition and noise control that Stable Diffusion users take for granted. Before you get excited, read the pack's own framing: these nodes are experimental, and HunyuanImage-3 is an autoregressive model, not a diffusion model. Latent manipulation behaves differently here than it does in SD, so results are less predictable and some things flat-out don't translate.

What it does specifically is take a HUNYUAN_LATENT tensor and transform it. Think of it as a filter rack for the noise/initial latent your generation starts from - you shape the starting point, and the model paints from your shape instead of from pure randomness.

How it works

It's a pure transform: HUNYUAN_LATENT in, HUNYUAN_LATENT out. The operation dropdown is the whole instrument, with seven modes:

• low_pass_filter - smooths the latent, keeping broad structure. Good for guiding macro composition (where subjects sit, general layout).
• high_pass_filter - keeps the fine detail, strips the broad layout.
• blend_seeds - mixes two noise seeds into one starting point.
• spatial_mask_left_right / spatial_mask_top_bottom - bias the latent across the frame.
• amplify_center - punch up the middle of the frame.
• invert - flip it.

Then two dials: strength (0–1, default 0.5) controls how hard the operation applies, and seed (default 42) controls the randomness for the seed-dependent operations like blend_seeds.

The HUNYUAN_LATENT type is its own thing - it comes from the pack's Hunyuan Empty Latent node (which builds a noise tensor at a specific resolution and seed), and it feeds the Hunyuan Generate with Latent node, which accepts it as the noise source instead of generating random noise itself. So the mini-workflow is: Empty Latent → Noise Shaping → Generate with Latent.

How to install it

Same 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

Results are unpredictable. This is the honest caveat and it's the model's fault, not yours. Because generation is autoregressive, the noise isn't a diffusion latent you can nudge the way you would in SD - the README explicitly warns img2img-style manipulation differs significantly, and recommends the composition mode (in Generate with Latent) as the more predictable path for layout guidance. Treat Noise Shaping as an exploration tool: throw a few strengths and seeds at it, see what sticks.

"The node needs a HUNYUAN_LATENT and I don't have one." Right - this node doesn't create noise, it shapes it. Start from Hunyuan Empty Latent and end at Hunyuan Generate with Latent. There's no standalone "encode" step; the pipeline is Empty Latent → (optional) Noise Shaping → Generate with Latent.

Still on the base model only. The latent family works with base (non-Instruct) models. The Instruct pipeline has its own generate nodes and doesn't take HUNYUAN_LATENT inputs - don't try to force the connection.