Y7 Prompt Enhancer (Native)

Most prompt enhancer nodes are little empires. They download their own LLM, load it through HuggingFace transformers, hold it in VRAM while your actual model waits, then fight ComfyUI over who unloads it. The Y7 Prompt Enhancer (Native) does the opposite: it takes a CLIP input from a stock CLIPLoader and makes the text encoder you already loaded rewrite your prompt. Nothing downloaded, nothing loaded, nothing left resident that ComfyUI isn't already managing. "Native" is the whole pitch.

This fits the 2026 model landscape better than it sounds. The newest image models don't read your prompt through CLIP anymore; they wrap it in a chat template and feed it to an actual LLM like Gemma or Qwen (HiDream-O1 ships Gemma-4 prompt enhancement in its stock workflows, LTX-2's text encoder is Gemma 3 12B). These encoders can not only encode text - they can generate it. This node cashes that in.

How it works

The mechanism is genuinely simple, and it's all on the clip object. The node checks the loaded encoder actually has a generate() path (more on that in a second), builds instruction + your text, then calls clip.tokenize(...) and clip.generate(...) - the same generation API ComfyUI's own Generate Text node uses. The output gets decoded, then a regex pass splits reasoning from answer.

The reasoning split is where the author earned their keep. Gemma 4 emits its planning inside a thought channel, and ComfyUI's decoder deliberately preserves that text - which is why the core Generate Text node hands you a wall of planning notes with your actual prompt buried at the end. This node separates the two, even in the awkward cases: a think block truncated by max_length, or Gemma reasoning past its primed channel and closing it with an orphan tag. Reasoning goes to thinking_output, the clean prompt to enhanced_prompt.

Not every encoder qualifies. Gemma 3/4, Qwen3, Qwen3.5, and Qwen3-VL have generation paths; T5, UMT5, CLIP-L/G, Gemma 2, and LLaMA-3.1 don't. The node detects this up front and tells you in plain English rather than blowing up somewhere deep inside the encoder. Nice touch.

The inputs that matter

• clip - the generation-capable encoder, loaded with a plain CLIPLoader.
• text - your rough prompt. Wire in any string node.
• instruction - the system-style instruction prepended to your text. This is the real dial; the default asks the model to "describe only what is visible" and keep every element you asked for. Editing it changes the whole flavor of the rewrite.
• max_length - maximum new tokens, default 2048. Not the context window. Two gotchas in one field: reasoning eats from this same budget, and the KV cache reserves roughly 84KB of VRAM per token up front.
• temperature, top_k, top_p - Gemma's recommended values are already set (1.0 / 64 / 0.95). temperature 0 switches to greedy decoding and ignores the other two.
• seed - identical inputs return a cached result, so bump this to re-roll.
• thinking - optional; lets the model reason first. Its output always goes to thinking_output, never into the prompt.

Wire enhanced_prompt into your conditioning (a CLIPTextEncode or the model's prompt input), and thinking_output into a Show Anything if you're curious what it was deliberating.

Install and models

ComfyUI Manager → Custom Nodes Manager → search Y7 or Y7Nodes, install, restart. Or the manual way:

cd ComfyUI/custom_nodes
git clone https://github.com/yushan777/ComfyUI-Y7Nodes
cd ComfyUI-Y7Nodes && pip install -r requirements.txt

The pack's requirements.txt needs transformers and sentencepiece - already present in a normal ComfyUI install, but worth knowing if you run a stripped-down venv. The actual model is the real download. Safetensors only, in models/text_encoders/ - e.g. Gemma 4 from Comfy-Org/gemma-4:

ComfyUI
└── models
    └── text_encoders
        └── gemma4_e4b_it_bf16.safetensors

GGUF will not work - ComfyUI core has no GGUF loader, and for Gemma 4 the tokenizer lives inside the safetensors file. For Gemma 4 the CLIPLoader type dropdown is ignored entirely; the model is detected from the weights.

Where people get burned

• enhanced_prompt comes back empty. The model spent the whole max_length budget reasoning. Raise max_length or drop temperature toward 0.
• "This text encoder does not support text generation." That's the guard doing its job - you fed it a T5 or CLIP encoder. Grab a Gemma/Qwen.
• Out of memory on generation. A big max_length on a 27B Qwen3.5 reserves a lot of KV cache. Start with the default and remember the ~84KB/token math.
• Don't confuse it with the LM Studio nodes in this same pack. Those need a running LM Studio server. This one needs no server, no key, nothing.

The author's README opens with "probably only useful to me. There's really nothing new here." That's the self-deprecation of someone who built exactly the enhancer they wanted: no model management, no 30-second first run while it downloads an LLM, just the encoder you're already paying for doing a useful second job.