Nodes/ComfyUI Easy Use/EasyLoader (HunyuanDiT)
ComfyUI Node Runs on cloud

EasyLoader (HunyuanDiT)

Load Tencent's HunyuanDiT in one node

By yolain·Created 3 years ago·Updated 8 days ago· 2,633
EasyLoader (HunyuanDiT)
  • optional_lora_stack
  • optional_controlnet_stack
  • pipe
  • model
  • vae
ckpt_name
vae_name
lora_name
lora_model_strength1.00
lora_clip_strength1.00
resolution1024 x 1024
empty_latent_width1024
empty_latent_height1024
positive
negative
batch_size1

HunyuanDiT is Tencent's diffusion-transformer text-to-image model, and its headline feature was bilingual prompting - it understands Chinese and English natively, which at its 2024 release was genuinely rare. easy hunyuanDiTLoader is yolain's one-node setup for it: pick the checkpoint, set your prompts and resolution, and out comes an EasyUse pipe ready for the rest of the chain.

It landed in ComfyUI-Easy-Use v1.2.0 next to the PixArt loader, both riding on city96's ExtraModels work for the DiT plumbing. Being honest about where this sits in 2026: HunyuanDiT is an older architecture that the frontier has largely moved past - Flux, SD3.5, Qwen and friends took the spotlight. But if you're specifically working with Hunyuan-DiT (bilingual prompts, a workflow that already uses it, or just curiosity about the DiT lineage), this loader is the tidy way in.

How it works

Structurally it mirrors the other EasyUse loaders. It loads the ckpt_name checkpoint, optionally applies a lora_name with its strengths, encodes positive/negative prompts, and builds an empty latent at your chosen resolution - the default here is a sensible 1024×1024, since DiT models like Hunyuan are trained for the higher resolution rather than SD1.5's 512. Everything bundles into the pipe output, with model and vae also broken out for direct use.

Because it's a transformer-based model rather than a UNet, there's no clip_skip knob here - that's a UNet-era CLIP setting that doesn't apply the same way. The loader keeps the interface deliberately close to comfyLoader so you don't have to relearn it.

The inputs that matter

  • ckpt_name - your HunyuanDiT checkpoint.
  • positive / negative - prompts; this is the model where you can genuinely write Chinese if you want.
  • resolution - a 1024-class preset, or custom width/height.
  • lora_name - optional Hunyuan LoRA, None otherwise.

Optional optional_lora_stack and optional_controlnet_stack inputs accept stacks built elsewhere in EasyUse, and the pack added ControlNet support for HunyuanDiT back in v1.2.2 if you want structural conditioning.

Installing it

The loader ships with the pack: ComfyUI Manager → "ComfyUI Easy Use" → install → restart. Or cd ComfyUI/custom_nodes && git clone https://github.com/yolain/ComfyUI-Easy-Use, run install.bat / pip install -r requirements.txt, restart. The catch is the model itself - HunyuanDiT's weights are a separate, sizeable download, and DiT loading leans on the ExtraModels support the pack integrates. Empty checkpoint dropdown means you haven't placed the model where ComfyUI can find it yet.

Common issues

The usual DiT-model friction. First, VRAM and download size - these are bigger, heavier models than an SD1.5 checkpoint, so expect a real download and don't be surprised by memory pressure on a small card. Second, HunyuanDiT expects Hunyuan-family LoRAs and ControlNets; an SDXL LoRA won't apply to it, and trying will error rather than silently do nothing.

And the strategic caveat worth saying plainly: this is not the model most people should reach for in 2026. If you don't have a specific reason to be on HunyuanDiT, a current checkpoint through easy comfyLoader (or a Flux/PixArt-Sigma loader) will treat you better. Use this loader when Hunyuan-DiT is genuinely the model you want.

CategoryEasyUse/Loaders

Inputs (13)

NameTypeDefaultDescription
ckpt_nameCOMBO0 options:
vae_nameCOMBO1 options: Baked VAE
lora_nameCOMBO1 options: None
lora_model_strengthFLOAT1.00-10–10
lora_clip_strengthFLOAT1.00-10–10
resolutionCOMBO1024 x 102431 options: width x height (custom), 512 x 512, 512 x 768, 576 x 1024, 768 x 512, 768 x 768, +25
empty_latent_widthINT102464–16384
empty_latent_heightINT102464–16384
positiveSTRING
negativeSTRING
batch_sizeINT11–64
optional_lora_stackoptLORA_STACK
optional_controlnet_stackoptCONTROL_NET_STACK

Outputs (3)

NameTypeDescription
pipePIPE_LINE
modelMODEL
vaeVAE