H3 CLIP LoRA

ComfyUI's core Load LoRA node demands a MODEL input. That's fine for a diffusion graph - but the H3 prompt-rewriter workflow has no diffusion model. It's just a text encoder feeding Generate Text, and you still want to slap the rewriter LoRA onto that encoder. H3 CLIP LoRA is the CLIP-side counterpart of LoraLoaderModelOnly: it applies a LoRA to a text encoder with no MODEL in sight.

How it works

Three inputs: clip (the text encoder the LoRA goes onto), lora_name (a LoRA from your loras folder), and strength (default 1.0, and the tooltip's range runs to ±100 for the brave). One output, clip, which you chain on - that's the whole contract.

Two things make it the right tool for this specific job. It's for text-encoder LoRAs keyed as text_encoders.transformer.* - the rewriter LoRA from lightx2v's MiniMax-H3-Prompt-Rewriter repo is exactly that shape, and so are the abliteration LoRAs people stack under their rewriters. And because it's a passthrough, you can chain two of them to stack LoRAs - the classic build being an abliteration LoRA under the rewriter LoRA, both on the same encoder.

The one warning

Keep the rewriter off any encoder that also builds conditioning. The rewriter LoRA changes how the model writes - it's for the prompt-generation pass, not for the CLIP that conditions H3 Encode. Using the same modified encoder for both is how you end up with a conditioning that's been rewritten in a voice it was never meant to speak in.

Install

Pack install: ComfyUI Manager → "Nynxz H3", or:

cd ComfyUI/custom_nodes
git clone https://github.com/Nynxz/ComfyUI-NynxzH3

Restart. Pure Python, no dependencies, ComfyUI 0.30.0+. The LoRA files themselves are separate downloads - grab ComfyUI-format versions from the nynxz/Qwen3-VL-8B-ComfyUI repo if you don't have them.

The one-liner

If your graph loads a text encoder but no model and you've ever stared at Load LoRA's red MODEL socket, this is the node. Chain two to stack, keep strength at 1.0 for the rewriter, and don't let it anywhere near the conditioning path.