Nucleus-Image Text Encode

This is the node that turns "a red fox sitting on a bench, golden hour" into the vector math Nucleus-Image's transformer actually reads. Nucleus-Image Text Encode encodes one prompt into a NUCLEUS_CONDITIONING - and it's also the moment the pack's lazy text-encoder loader finally does its job.

How it works

The node takes the NUCLEUS_TE handle from Text Encoder Loader (which, remember, loaded nothing), then on this execution it actually builds the Qwen3-VL encoder on the GPU, loads the FP8 weights, and encodes. Mechanically it:

1. Wraps your prompt in the pack's system instruction - the one about following the prompt literally, keeping left things left, and matching counts and colors.
2. Applies the Qwen chat template and tokenizes (capped at 1024 tokens, truncating the rest).
3. Runs the encoder and grabs hidden state -8 - the second-to-last-ish layer - as the prompt embedding, plus an attention mask.
4. Frees the encoder from the GPU and empties the cache.

The load-encode-release cycle is the whole point. It costs you a few seconds per encode, but it means an 8.2GB encoder doesn't sit in VRAM while the transformer wants it. If that churn annoys you, the Dual variant is the fix.

Inputs and output

• text_encoder - the NUCLEUS_TE output from the Text Encoder Loader.
• text - multiline prompt box. Empty string is legal and useful: the README explicitly says an empty "" works as a negative condition.

Output is a single NUCLEUS_CONDITIONING, which plugs into the Sampler's positive or negative input. For a full run you'll need two of these nodes - one positive, one negative - or one Dual node.

Prompt style, because it actually changes output

Since the encoder is an LLM, write like you're instructing it: full sentences, explicit layout, exact colors and counts. "A man in a red jacket standing on the left of a brick wall, three pigeons on the ground to his right" will land far better than "man, red jacket, brick wall, pigeons". This is the same shift the KB documents for every Qwen-encoded model - CLIP-era habits (tag soup, weight syntax like (word:1.2)) don't carry over. What also doesn't carry over: any negative embeddings you've collected for SDXL.

Installing and troubleshooting

Pack install is standard - ComfyUI Manager (search "Nucleus-Image") or git clone https://github.com/a180265/Nucleus-Image-comfyui-beta into custom_nodes/, then restart. The file this node needs is nucleus_image_text_encoder_fp8.safetensors in models/text_encoders/. No tokenizer to fetch; it ships with the pack.

The failure you'll most likely hit is the README's FAQ: images unrelated to the prompt. The author's fix is almost always "your negative prompt isn't wired up" - make sure a negative conditioning is connected to the Sampler, even if it's an empty string. If the node itself errors, nine times out of ten the encoder file is missing or in the wrong folder.

One more honest beta note: everything in this pack was verified on FP8. If you picked the bf16 encoder file, you're on an untested path - fine for experimenting, just don't blame the pack if something smells off.