Hunyuan DiT Text Encode (simple)

Hunyuan DiT had a real moment. Tencent dropped it in mid-2024 as an open, SD3-style diffusion transformer, and the launch thread ("JUST out - open source SD3-like architecture text-to-image model") pulled over 350 upvotes and 200+ comments - right when the community was actively hunting for something better than SD3's disappointing release. It got Diffusers support fast, kohya and webui integration was announced within weeks. Then it caught an EU compliance scare that spooked would-be adopters right as Flux landed and ate all the remaining oxygen. Tencent's next model, Hunyuan Video, went on to be huge - but the image model stayed a footnote. So treat this less as "join a thriving ecosystem" and more as "run a genuinely interesting architecture nobody's actively building on anymore."

This node is your text-conditioning entry point for it, and it's the "simple" version on purpose.

What "simple" means here

Hunyuan DiT conditions on two text encoders at once, concatenated - a Chinese RoBERTa-based CLIP and mT5 (multilingual T5) - the same dual-encoder instinct Flux later used with CLIP+T5. The full encode node in this pack lets you send different text to each branch. This "simple" version collapses that down to one textbox: whatever you type goes to both encoders identically. Unless you specifically want to steer the two encoders apart, this is the one to reach for - it's one less thing to think about.

Inputs and output

• text - your prompt, plain multiline string. Both English and Chinese work; the Chinese RoBERTa branch was trained bilingually.
• CLIP - the Chinese-RoBERTa-based text encoder, loaded from Hunyuan DiT's own CLIP loader in this pack.
• T5 - the mT5 encoder, loaded via the T5v1.1 Loader node pointed at Hunyuan's mT5 weights (not the DeepFloyd T5-XXL used elsewhere in this pack - different file).

Output is a single CONDITIONING - wire it straight into your KSampler's positive and negative inputs, same as any other conditioning.

Installing it

Part of the whole ComfyUI_ExtraModels pack:

• ComfyUI Manager - search "Extra Models for ComfyUI", install, restart.
• Manual - cd ComfyUI/custom_nodes && git clone https://github.com/city96/ComfyUI_ExtraModels, then pip install -r requirements.txt inside your venv. Restart.

Then the Hunyuan-specific downloads, all from Tencent's HF repo:

• First text encoder → ComfyUI/models/clip, renamed to chinese-roberta-wwm-ext-large.bin
• Second text encoder (mT5) → ComfyUI/models/t5, renamed to mT5-xl.bin (or the smaller fp16-converted version if you want faster loading)
• The Hunyuan checkpoint itself → ComfyUI/checkpoints, renamed to HunYuanDiT.pt
• Any SDXL-class VAE

Common issues

This is explicitly marked WIP in the pack's own README, so expect rough edges rather than a polished product. The two encoder downloads are large and easy to mismatch - double-check you renamed the CLIP file and the T5 file into the correct folders (models/clip vs models/t5), since a misplaced file just silently fails to appear in the loader dropdowns rather than throwing an obvious error. Use an SDXL VAE, not an SD1.5 one - Hunyuan DiT's latent space matches SDXL's, and the wrong VAE gives you visibly wrong colors and structure, not a crash.

Given how thin ongoing community attention is here, if you hit something that looks like a real bug rather than a setup mistake, the pack's GitHub issues are a better bet than searching for a fix on Reddit - most of that discussion moved on to Hunyuan Video years ago.