Replicate lucataco/hunyuandit-v1.1

HunyuanDiT was Tencent's first open text-to-image model - an SD3-class diffusion transformer released in May 2024, with one real distinction: it's genuinely bilingual, trained to understand both Chinese and English prompts. That's the reason to reach for it. If you're prompting in Chinese, or want a model that natively handles Chinese cultural concepts rather than mangling them, HunyuanDiT is a natural fit where most Western models struggle. This node runs the v1.1 checkpoint through Replicate - REPLICATE_API_TOKEN, pay-per-run, nothing downloaded.

Straight talk on its standing: HunyuanDiT landed weeks before Flux and got steamrolled. It's a competent, largely forgotten model in the English-speaking community. But "forgotten" isn't "bad," and for bilingual work it still has a clear niche.

How it works

It's a DiT (diffusion transformer) in the same family lineage as SD3, wrapped by lucataco - one of Replicate's most prolific packagers, the person who's put hundreds of open models onto the platform. The node relays your prompt and settings to the hosted model; a GPU on Replicate's side does the diffusion; an image returns. Nothing about your local hardware matters.

The inputs and outputs that matter

• prompt (default "a cute cat") - write it in Chinese or English; that's the model's party trick.
• negative_prompt - what to steer away from. HunyuanDiT is a full CFG model, so negatives actually work here (unlike distilled models where they don't).
• size (square / landscape / portrait) - simple shape control.
• num_inference_steps (40) - quality vs time; 40 is a generous default.
• guidance_scale (6) - prompt adherence, standard-diffusion range.
• sampler (ddpm / ddim / dpmms) - the sampling method; ddpm is the default and fine to leave.

enhance_prompt (a boolean auto-embellisher) and seed round it out; force_rerun forces a fresh run. Output is a single IMAGE.

How to install it

One install covers every Replicate node.

• ComfyUI Manager: search ComfyUI-Replicate, install, restart.
• Manual:

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

  then restart.

Nothing lands on your GPU - the model runs on Replicate's servers, so the only install is the replicate Python client. Set your token before launching ComfyUI: export REPLICATE_API_TOKEN="r8_..." (replicate.com/account/api-tokens). No token, no runs.

Common issues & troubleshooting

Auth error first run. REPLICATE_API_TOKEN needs to be set before ComfyUI starts. Set it, restart.

Why use this over Flux or a newer model? For English-only work, honestly you probably wouldn't - newer models are stronger. HunyuanDiT's edge is bilingual/Chinese prompting. Reach for it when that's the job.

Anatomy and fine detail look dated. It's a 2024 model that never got the finetune ecosystem later bases enjoyed. That's the ceiling - no setting fixes it.

Same image on re-run. Replicate returned a cached result for identical inputs. Change the seed or set force_rerun.

It bills per run. Each generation is a paid Replicate prediction. Standard cloud-model economics - cheap per image, but not free.