Hunyuan 3 Generate (Large/Offload)

The standard Hunyuan 3 Generate node is happy up to about 2MP. Want a 4K render, a 6MP landscape, an 8MP poster? That's what Hunyuan 3 Generate (Large/Offload) (class HunyuanImage3GenerateLarge) is for - a generate node with a real resolution dropdown (2K, 4K UHD, 6MP all listed) and an offload_mode that will move the model's weights to CPU RAM when VRAM runs out mid-generation. It's the "I have the RAM, let's go big" node.

Two honest caveats before you build a workflow on it. First, it's for BF16 (full-precision) models - the README's compatibility matrix is explicit that the Large/Offload path conflicts with quantized NF4/INT8 weights, which can't be moved to CPU correctly. If you're on NF4, use the Low VRAM generate variants instead. Second, it OOMs at 3MP+ even on 96GB cards - that's the MoE dispatch_mask problem, and the fix is the pack's HighRes Efficient node. Large/Offload owns the 2MP–3MP-with-offload zone; beyond that, switch.

The inputs that matter

Mostly standard generate fare - model (HUNYUAN_MODEL), prompt, seed, steps (40), guidance_scale (7.5) - plus the two that define this node:

• resolution - 20 options from 720p up through 2560x1440 2K (3.7MP), 3840x2160 4K UHD (8.3MP), and 3072x2048 6MP (6.3MP). Here's the subtle bit from the README: picking "Auto (model default)" makes the node assume ~2.5MP to be safe, which can trigger CPU offload even for a small image. Pick the specific resolution and it knows exactly how much VRAM is needed and skips offload when you fit.
• offload_mode - smart (default, offloads only when VRAM is critically low), always (move to CPU first, slow but safe), disabled (force everything on GPU, fastest, OOM if you misjudge).

The post_action dropdown (keep_loaded / soft_unload_to_cpu / full_unload) controls what happens after generation. The optional prompt-rewrite cluster (enable_prompt_rewrite, rewrite_style, api_url, model_name) needs an OpenAI-compatible LLM API key - that's optional and off by default.

Outputs: image, rewritten_prompt (STRING, only populated if rewriting is on), status (STRING), and a trigger passthrough for sequencing.

How to install it

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

Restart ComfyUI, hard-refresh the browser, or search HunyuanImage in ComfyUI Manager. Pair it with the Full BF16 Loader - the README's recommended combo for big BF16 canvases.

Common issues & troubleshooting

Slow generation = you're offloading when you don't need to. The README's performance guide is basically this node's user manual. Reserve enough VRAM in the loader (reserve_memory_gb 30GB+ for 4K) so Smart mode doesn't panic-offload, and always pick a specific resolution instead of Auto.

OOM at 3MP+. Expected on the standard path - switch to Hunyuan 3 Generate (HighRes Efficient), which replaces the memory-hungry MoE dispatch with a loop-based approach (~75× less intermediate VRAM) and is the node actually built for 3MP–4K.

It's suddenly in CPU-speed mode with NF4. Wrong node for quantized models. Swap to the Low VRAM generate variants, which skip the conflicting CPU-offload calls.