MiaoBi CLIP Loader

MiaoBi (妙笔) is a Chinese-language text-to-image checkpoint from developer ShineChen1024, built on the exact same architecture and latent space as SD1.5. That's the whole pitch: it's a drop-in SD1.5 sibling trained on Chinese captions instead of English or machine-translated data, so it plays nice with your existing SD1.5 LoRAs, ControlNets, and T2I-Adapters - you're not stepping into a separate ecosystem, just a differently-trained text encoder and UNet pair.

Worth setting expectations honestly here: this is a niche model. It has almost no visible footprint under its own name in the wider ComfyUI/SD community, and it's easy to confuse with two totally unrelated things that share part of the name - "miaobishenghua" is a popular but unrelated LoRA style, and "MiaoMiao" checkpoints are a different model family entirely. If you go looking for troubleshooting threads and find one of those instead, you're not in the right place.

This node loads MiaoBi's text encoder specifically - the CLIP-equivalent half of the pair.

What it does

It's a standard CLIP-style loader: reads a safetensors checkpoint off disk and hands back a CLIP object in the exact shape ComfyUI's normal text-encode and sampling nodes already expect. Nothing MiaoBi-specific happens inside the loader itself - the model-specific part is entirely in what weights you feed it.

• clip_name - a dropdown populated from files sitting in ComfyUI/models/clip. Point it at MiaoBi's converted CLIP file.
• Output: CLIP - feed into a normal CLIP Text Encode node, prompting in Chinese for best results since that's what the model was actually trained on.

Pair this with a separate checkpoint/UNet loader pointed at MiaoBi's diffusion weights to get a full pipeline - or skip both and use MiaoBi Checkpoint Loader (Diffusers) instead if you downloaded the whole Hugging Face repo rather than hand-picking files.

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. Restart.

Then, from MiaoBi's own Hugging Face repo:

• Download text_encoder/model.safetensors, rename it to something clearly identifiable like MiaoBi_CLIP.safetensors, and drop it in ComfyUI/models/clip.
• Download unet/diffusion_pytorch_model.safetensors, rename to MiaoBi.safetensors, and put it in ComfyUI/models/unet for the matching model loader.

Common issues

The biggest trap is silent, not loud: this loader doesn't validate that the file you picked is actually MiaoBi's encoder. Point it at any unrelated SD1.5 CLIP file and it'll load fine and generate images - just not MiaoBi's outputs, and with no error telling you why results look off. Keep your filename unambiguous so future-you doesn't lose track of which CLIP is which.

Prompting in English will still technically work since the underlying architecture is SD1.5-shaped, but you're leaving the model's actual training advantage on the table - this thing exists specifically because Chinese-language prompting on Western SD1.5 checkpoints was mediocre. And because community adoption is thin, don't count on finding a Reddit thread for whatever you run into; the pack's own GitHub issues are the more realistic first stop.