Hunyuan Instruct Multi-Image Fusion

"Hunyuan Instruct Multi-Image Fusion" is the pack's flashiest Instruct node: it takes up to five reference images and combines them into one output based on a plain-language instruction. "Put the cat from image 1 in the scene from image 2 with the lighting from image 3" isn't a hypothetical - that's the README's example, and it works. Where Image Edit changes one image, Fusion mixes several, and it's one of those capabilities that makes HunyuanImage-3.0's Instruct variant feel like it's from a different generation than the diffusion stack.

The inputs that matter

• image_1 through image_5 - the references. image_1 is required; the rest are optional. The model officially supports three; slots 4 and 5 are experimental - the tooltip is upfront that they increase VRAM significantly and results vary.
• instruction - how to combine them, referencing "image 1", "image 2", etc. English or Chinese both work.
• bot_task - image (direct fusion), recaption (rewrite the instruction into a detailed fusion plan), or think_recaption (CoT reasoning about how to combine elements, then rewrite, then generate). The tooltip recommends think_recaption for complex fusion. Remember the cost: CoT text generation happens before any image generation.
• resolution - default 1024x1024. The tooltip's warning is worth quoting: with multiple differently-sized inputs, "Auto lets the model predict aspect ratio (often unpredictable)"; picking a specific preset gives deterministic dimensions.
• system_prompt - dynamic by default; en_unified adds reference-editing guidelines for extracting and combining elements across images.
• steps / guidance_scale / flow_shift / max_new_tokens - same auto semantics as the Image Edit node. max_new_tokens at 3072 is suggested for complex fusion instructions.

Outputs: image, cot_reasoning (the model's combination plan, as a STRING), and status.

Workflow shape

Load Image 1 ─┐
Load Image 2 ─┼→ Hunyuan Instruct Multi-Fusion → Save Image
Load Image 3 ─┘

The model input plugs into a Hunyuan Instruct Loader. The think_recaption mode plus matching input resolutions is the recipe the README uses in its example workflow - mismatched input sizes are exactly when the "unpredictable aspect ratio" warning bites.

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"). You need an Instruct model - Distil-NF4 for 48GB cards, Distil-INT8 for 96GB:

cd ComfyUI/models
huggingface-cli download EricRollei/HunyuanImage-3.0-Instruct-Distil-NF4-v2 --local-dir HunyuanImage-3.0-Instruct-Distil-NF4-v2

The honest caveats

Fusion with 2–3 images is a headline feature of the Instruct model; 4–5 images is the pack author pushing beyond what the model was trained to do, and the tooltips and README say so explicitly. Expect the results to degrade and the VRAM to spike as you add slots. And this is the same slow territory as the other Instruct nodes: an 80B model reading several images and reasoning about them takes minutes. If your fusion output looks like the model just blended styles randomly, the fix is usually a more specific instruction and think_recaption, not more slots.