Hunyuan Generate with Latent

"Hunyuan Generate with Latent" is the experimental node that adds image and latent inputs to the Unified Generate V2 - it inherits all of V2's auto-detection, block swap, and VRAM management, then layers on two optional inputs: an image (for image-guided generation) and a latent (for custom noise injection). It's the pack's answer to the thing everyone wants from an image model: composition control and img2img.

The README is careful to set expectations up front: HunyuanImage-3.0 is autoregressive, not diffusion, so latent manipulation doesn't behave the way it does with Stable Diffusion. Don't expect faithful SD-style img2img. That caveat is the whole article in miniature - this node is genuinely useful, but it's a different beast.

The modes

When an image is connected, image_mode decides how it influences generation:

• composition (default) - extracts broad spatial layout via a low-pass filter in latent space and modulates noise amplitude. No ghosting; only macro composition transfers. This is the mode the README recommends for predictable layout guidance.
• img2img - the traditional latent mix: (1−σ)·clean + σ·noise. Low denoise preserves the image, high denoise adds variation - and low denoise can ghost.
• energy_map - uses per-channel energy (absolute magnitude) to scale noise spatially. Abstract and artsy.

denoise_strength (default 0.6) means different things per mode: for composition/energy_map it's how strongly the layout modulates noise; for img2img it's how much noise replaces the image.

Usage combinations: image + no latent = image-guided generation; latent + no image = custom noise base; both = image-guided with custom noise; neither = plain V2 text-to-image.

The latent input

The latent input takes a HUNYUAN_LATENT tensor from the pack's own Hunyuan Empty Latent node (random noise at a given resolution/seed) or Hunyuan Latent Noise Shaping (frequency-filter, amplify, invert). That's the injection point for the experimental latent-control workflow - same seed and resolution as the V2 node should reproduce its output exactly, which doubles as a sanity check that your latent is actually being used.

Everything else matches Unified V2: model_name, prompt, resolution, num_inference_steps, guidance_scale, seed, plus blocks_to_swap, vae_placement, post_action, enable_vae_tiling, flow_shift, reserve_vram_gb, moe_drop_tokens, vae_dtype, and force_reload. Outputs are images and final_prompt.

Installing

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

Restart ComfyUI (or ComfyUI Manager, search "Comfy_HunyuanImage3"), and have a base (non-Instruct) model in ComfyUI/models/:

huggingface-cli download EricRollei/HunyuanImage-3-NF4-v2 --local-dir HunyuanImage-3-NF4

The honest read

This is marked experimental for a reason - the author's own docs flag it as behavior that differs significantly from diffusion img2img, and the latent-control nodes shipped in v1.3.0 are the newest code in the pack. If you want predictable layout guidance, use composition; if you want classic img2img, you'll likely be frustrated. And because it inherits V2's auto-magic, pair it with the same warnings as Unified V2: on a 24GB card the split Low VRAM loader/generator pair is still the more reliable path.