Nucleus-Image Text Encode (Dual)

If you're running Nucleus-Image properly, you need a positive and a negative prompt - and this is the node the pack's README pushes you toward for exactly that. Nucleus-Image Text Encode (Dual) encodes both in a single visit to the Qwen3-VL text encoder, instead of the two round-trips you'd get with a pair of plain Text Encode nodes.

Why it exists

Remember the lazy-loading scheme: the Text Encoder Loader loads nothing, and each Text Encode node hauls the whole 8.2GB encoder onto the GPU, encodes, and throws it away. Do that twice (once for positive, once for negative) and you've paid the load cost twice - a few extra seconds of VRAM churn every run. This node loads the encoder once, runs both prompts back to back, and releases it. The author's own comment in the source calls it "roughly halves encoding time," and on a model this size that's not nothing.

The inputs and outputs

• text_encoder - NUCLEUS_TE from the Text Encoder Loader.
• positive_text - the main prompt. Multiline.
• negative_text - the negative prompt. Multiline, and an empty string is fine.

It returns two outputs, both NUCLEUS_CONDITIONING: positive and negative, which wire straight into the Sampler's matching inputs. That's the whole reason to prefer this over two single-encode nodes - one node, both wires filled, one encoder load.

Actually getting good results from it

The negative prompt is not decorative on this model. The README's most common complaint - "my image has nothing to do with my prompt" - is answered with "make sure a negative is connected." A Dual node makes it trivial to check you've got one. And because the encoder is an LLM, write both sides as sentences, not tag soup: "no text, no watermark, no distorted hands" beats text, watermark, bad hands. The KB's Qwen-encoder guidance applies verbatim - full sentences with explicit ownership land better than comma lists.

Installing and troubleshooting

Pack-wide install, same as every node here: ComfyUI Manager (search "Nucleus-Image") or

cd ComfyUI/custom_nodes
git clone https://github.com/a180265/Nucleus-Image-comfyui-beta

restart, then make sure nucleus_image_text_encoder_fp8.safetensors is sitting in models/text_encoders/. The tokenizer and chat template are bundled in the pack, so there's no extra download.

Trouble-shooting is short because the node is simple. If it errors, it's the encoder file: missing, wrong folder, or pointed at a non-Qwen file. If the output looks unguided, your negative prompt may be empty and your cfg left low - see the Sampler article on how cfg below 1.0 stops guidance from using the negative at all. And one beta reality check: the whole pack was validated on the FP8 encoder, so stick to that file unless you're deliberately experimenting.