Hunyuan Unified Generate V2

"Hunyuan Unified Generate V2" is the pack's attempt to collapse the whole loader-plus-generator dance into a single node. It has a built-in loader, auto-detects your model's quant type from the folder name (NF4/INT8/BF16), handles block swap, memory budgets, and VRAM management, and outputs an image. No separate loader node, no HUNYUAN_MODEL wiring, no decisions about which generate variant matches which loader. You pick a model folder and a prompt, and it figures the rest out.

It's a genuinely good idea, and it's also the node the README positions as the base-model replacement: "Single node replaces all base-model generate variants." The catch is the same one every all-in-one node has - you trade the explicit control of the split loaders for convenience, and when something goes wrong, the one node is a black box.

What you set

The required inputs are the essentials: model_name (dropdown of detected model folders), prompt, resolution (40 presets at common photo ratios), num_inference_steps (default 40; the tooltip notes 50–80 reduces flow-matching artifacts at 2K+), guidance_scale (default 5.0), and seed (-1 for random).

The optional inputs are where the unified node shows off:

• blocks_to_swap - default 20, with -1 meaning auto-calculate. 0 = no swapping (NF4 needs ~50GB; BF16 falls back to device_map).
• vae_placement - auto (decide from VRAM), always_gpu, or managed (VAE moves to CPU when not decoding).
• post_action - full_unload by default here, unlike the split generators' keep_loaded. Worth knowing if you expected the model to stay warm.
• enable_vae_tiling - tile the VAE decode for large images (slower, less VRAM).
• flow_shift - default 2.8; lower for portraits, higher for landscapes.
• reserve_vram_gb - set aside VRAM for downstream nodes (upscalers, other models).
• moe_drop_tokens - default true; false for best quality on ≥48GB cards.
• vae_dtype - bfloat16 default, float32 to cut banding.
• force_reload - for orphaned VRAM after a failed load.

Outputs are images (an IMAGE) and final_prompt (a STRING - handy, since there's no rewritten-prompt/status trio here).

When to use it

For a clean text-to-image graph that you don't want to maintain, this is the node. It's also the base that Hunyuan Generate with Latent inherits from - the experimental latent-control node is literally a subclass of this one, so anything you learn here transfers.

Installing

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

Restart ComfyUI, or install via ComfyUI Manager (search "Comfy_HunyuanImage3"), then drop a model in ComfyUI/models/ - e.g. huggingface-cli download EricRollei/HunyuanImage-3-NF4-v2 --local-dir HunyuanImage-3-NF4.

When not to

If you're on a 24GB card, the split NF4 Low VRAM+ loader → Low VRAM Budget generator pair is still the recommended road - that pairing's device_map strategy was specifically rebuilt to avoid the bitsandbytes validation errors that plagued early NF4 attempts, and the unified node's auto-magic can't replicate the fine-grained budget control. And for 3MP+ BF16 renders on 96GB, the HighRes Efficient generator is the specialist that this node doesn't try to be. Unified V2 is the great default; the split nodes are the specialists.