Gemma Prompt Expander Native (Tags / Natural Language)

Most prompt-enhancer nodes give you one string and hope you like it. This one gives you two, from the same scene, and lets you decide which one your model actually wanted. That's the whole reason it exists - and it's a sharper design than it sounds.

What you're actually getting

Gemma Prompt Expander Native sits in the prompt category of RS-Seasonal-Prompt-Generator, a nine-node pack from a solo Korean author (raspie) who writes every doc twice, in Korean and English. The pack's nodes are all variations on one idea: use a small Gemma as an in-graph writer, either to turn your rough idea or a reference image into Danbooru tags, or into a plain-English instruction sentence.

Why you'd reach for it: tag models and prose models want opposite things. Illustrious, NoobAI and Anima were trained on the Danbooru vocabulary and reward 1girl, solo, dutch angle over a nicely written paragraph; Flux, Z-Image and Klein read prose and quietly degrade on comma soup. Having both outputs means one node can feed either kind of graph, and you can compare on the same seeded scene instead of guessing. The broader split is covered in the Danbooru tags panel and it's the single most common way people mis-wire a prompt node.

I'll be blunt about the reputation side: there isn't one. Search the community corpora for "Gemma TIPO" or the author's handle and you get nothing - this pack has no Reddit footprint at all. What it does have is documentation that's more honest than most popular packs: the author leads with "outputs may be wrong or not match your taste," and says outright that Korean edits, names and appearance stay unreliable. Take the warnings at face value and you'll have a decent time.

How the two outputs get made

output_format is the switch, and the mechanism matters:

• both (default) runs two sequential calls on the same model - first a prose scene, then a second pass that renders that exact scene as tags. The image, if you connected one, is encoded once and not re-encoded for the second call. You load one model, not two, but you pay twice the wall time.
• tags runs one call and leaves natural_language as an empty string.
• natural_language runs one call and leaves tags empty. It asks for a Create … image instruction, not a description and not a tag list.

Both output languages are English regardless of what you type. Korean, English or Japanese input goes to the model verbatim, and tag post-processing (dedup, ban filtering, conflicting person-count cleanup) is only applied to the tag path - prose is left alone.

The inputs you'll actually touch

Seven required widgets, and honestly most people set three of them.

• prompt - what to keep, change, or drop. A line like "keep the blue hair, put her at a winter night market holding tea, coat on, no scarf" works better than a keyword pile.
• output_format - see above. Start with tags to judge the model cheaply, then try both.
• tag_length - very_short … very_long, a detail target, not a count. Defaults to long.
• max_tokens - per call, default 512. With both you're spending two of these.
• ban_tags - comma-separated English exclusions. The author's tooltip is precise and worth believing: exact filtering applies to tags only; prose just follows the instruction. It won't reliably remove something from a sentence.
• temperature / seed - 0.5 default. Keep temperature positive: current ComfyUI ignores repetition penalties at temperature=0, so the repetition_penalty widget (tag call only, prose stays at 1.05) does nothing at zero.

On the optional side, only three earn their space for a beginner. model_source picks your base - connected_clip (a native Gemma 4 E2B from Load CLIP) or heretic_ara_q4 (downloads its own base on first run). image adds one reference picture, encoded once. And tipo_lora offers three experimental adapters (pilot-001, pilot-002, image-30k) on top of none; pair it with tipo_lora_scope=tags_only so prose stays on the clean base. A LoRA alone cannot run - it needs one of the two bases.

Then the three outputs: tags, natural_language, and raw_output. Wire tags into a tag model's positive prompt, natural_language into a prose model's, and keep raw_output on a text preview when something looks wrong - it's the unfiltered model output, which is how you tell "the model didn't say it" from "the post-processor ate it."

Install

Manager first: search the pack title, install, restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/raspie10032/RS-Seasonal-Prompt-Generator

Watch the folder nesting - the author calls this out because it's the usual reason nodes don't appear; you want custom_nodes/RS-Seasonal-Prompt-Generator/__init__.py, not a nested copy.

The built-in Q4 path needs exactly one extra package, and it's tiny compared to what you'd expect:

python -m pip install -r ComfyUI/custom_nodes/RS-Seasonal-Prompt-Generator/requirements.txt

That's gguf>=0.19.0,<0.20, a pure-Python GGUF parser. No llama.cpp, no CUDA wheels, no torch replacement. Then disconnect clip, set model_source=heretic_ara_q4, leave the path fields blank, and the first run pulls a ~3.19 GiB text GGUF plus a 30.7 MiB tokenizer into ComfyUI/models/text_encoders/rs_gemma_q4/. Add an image and it fetches a ~0.92 GiB FP16 mmproj too. Files are SHA-256 checked against pinned revisions, and partial downloads are discarded rather than kept.

The connected_clip path skips all of that but wants a ComfyUI-format Gemma 4 E2B safetensors (~9.57 GiB) in models/text_encoders/ and a ComfyUI new enough to support Gemma 4 CLIP.generate.

Where people get burned

Two calls isn't free. A both run at 512 tokens is roughly twice the latency of a tags run. If you're iterating on a prompt, work in tags mode.

Connect a VRS state or a state file, never both. vrs_state (from VRS Native Image Recall) and vrs_state_path are mutually exclusive and the node throws if you wire both. If you use a state with a native query in it and also connect an image, the pixel hashes have to match - that's the node refusing to expand a scene you retrieved with a different picture.

The e621 path is a dead end with VRS. tag_set offers danbooru and e621, but an existing VRS generation state is Danbooru-only and gets rejected. Same for a tag dictionary that has no entries for your chosen set.

Silent blank output. If input validation fails, the node returns empty strings for all three ports rather than erroring loudly. If you get an empty prompt downstream and no red box on the node, check the validation path before you blame the model.

Disk size is not VRAM. The author says this in three separate places for a reason. ~3.22 GiB of files still needs restore buffers, image processing, KV cache and possibly a LoRA on top, plus whatever diffusion model is loaded. Tested on 16 GB NVIDIA cards; a 24-image eval on a 5060 Ti had a median call time around 36 seconds, ranging 16–126.

What I'd actually run

model_source=heretic_ara_q4, output_format=tags, temperature=0.5, max_tokens=512, seed fixed so you can compare, and repetition_penalty=1.15 only if you're drowning in repeats. Add both once you trust it, and wire each port to the model family that matches. It's a niche node from an unknown pack, but the dual-output design is the sort of thing bigger packs should have copied.