SP_HunyuanLoader

Hunyuan Video is not a small model - Tencent's December 2024 13B video generator wants 45–60GB of VRAM at fp16, which is why in practice everyone runs it quantized and why a loader that bundles the quant choices into one widget panel is actually convenient. SP_HunyuanLoader from the SP-Nodes pack is exactly that: one node that loads the Hunyuan unet, its two text encoders, the VAE, encodes your prompt, and builds the video latent - text-to-video or image-to-video - in a single pass.

Like the pack's other loaders it's a group node, expanded at runtime via ComfyUI's GraphBuilder into the stock nodes you'd otherwise place by hand: UNETLoader (or GGUF/bitsandbytes variants), DualCLIPLoader, VAELoader, CLIPTextEncode, and EmptyHunyuanLatentVideo. Its outputs are sp_pipe, model, clip, vae, positive, negative, latent, designed to feed the pack's own SP_KSampler through the custom SP_PIPE type, though the individual outputs drop into a stock sampler fine too.

The inputs that matter:

• unet_name + weight_dtype - same ladder as SP_FluxLoader: fp8 variants, nf4-float8_* (needs the author's bitsandbytes NF4 loader), or gguf (needs ComfyUI-GGUF). For a 13B video model on a consumer card, gguf or nf4 is usually the difference between running and not.
• hunyuan_fast - a boolean that switches the model shift applied via ModelSamplingSD3 from 7.0 to 17.0. Higher shift = fewer steps to a clean result; it's the community "fast sampling" trick for this model baked into the loader.
• empty_latent_width / height / length - defaults 848 × 480 × 73 frames, the canonical Hunyuan output. Length is the number of frames in your clip; the node caps it at 100.
• image, image_megapixels, image_tile_size - the image-to-video path. Feed a start image and the node scales it to your target megapixels (0 = skip), resizes to a multiple of 16, and VAE-encodes it - tiled if you set image_tile_size, which is your VRAM safety valve. Leave image unplugged for pure text-to-video.

Installing is the pack-wide story:

cd ComfyUI/custom_nodes
git clone https://github.com/bananasss00/ComfyUI-SP-Nodes

or ComfyUI Manager → SP-Nodes, then restart. The pack's own dependency table lists three extra packs for this node: ComfyUI-GGUF, ComfyUI_bitsandbytes_NF4-Lora, and ComfyUI_essentials (the essentials pack supplies the ImageResize+ and ImageScaleToTotalPixels+ nodes the image path calls). Don't skip them - if you pick a gguf dtype without ComfyUI-GGUF, or plug in an image without essentials, the node dies at expansion with a missing-node error, not a friendly warning.

One honest caveat: Hunyuan was the first mover in local video and then got left behind - by mid-2026 it's Wan and LTX territory for most people, and the KB tracks local usage collapsing. It's still the right call if you specifically want its cinematic look, or if you already have the weights. Start at 848×480 and 73 frames, keep image_tile_size on if VRAM is tight, and if 13B is out of reach on your card, note that this loader also accepts HunyuanVideo-1.5 unets (the 8.3B rebuild that fits in ~14GB) - just respect that it's a different model with different step and shift expectations.