RS-Seasonal-Prompt-Generator

RS-Seasonal-Prompt-Generator

English | 한국어

Generates season-specific fashion prompts by randomly combining fashion items, background settings, weather, time, and additional situational details from CSV data.

v2.7.0 merged the text and vision TIPO nodes into ONE node (image is an optional input — connect an image to use vision mode, otherwise it runs text mode) and added a simple sort preset. v2.7.1 makes vision mode fully automatic on all platforms (the llama.cpp mtmd binary now auto-downloads on Windows / Linux / macOS, not just Windows). v2.7.3 upgrades the auto-download default to the v2 Korean TIPO fine-tune (gemma4-tipo-ko-v2-Q4_K_M.gguf, better color/attribute accuracy). See the changelog.

This pack provides two independent nodes (category: prompt):

1. Seasonal Fashion Prompt Generator (CSV, zero-dep), 2) Gemma TIPO (Prompt / Image → Tags). Install by cloning/copying this folder into ComfyUI/custom_nodes/ and restarting ComfyUI. Each outputs a STRING; they're chainable.

Showcase

Korean natural language in → clean English Danbooru tags out:

In: 소녀가 하얀 세일러복과 스카프를 하고 있음.

Out: white sailor collar, white skirt, black hairband, blue eyes, blush stickers, long sleeves, looking at viewer, open mouth, red scarf, school uniform, serafuku, sitting, solo focus, sparkle sticker, striped clothes, sweatdrop, collared shirt, white background

And vision mode — connect an image, get a plausibly-expanded Danbooru tag set (TIPO-style; not exact labeling):

Node 1 — Seasonal Fashion Prompt Generator

The original generator. Pick a season and toggle the fashion / background / weather / pose / situation options; it returns a comma-separated prompt randomly combined from the bundled CSV data. Zero dependencies — nothing is installed and nothing runs a model.

Node 2 — Gemma TIPO (Prompt / Image → Tags)

One unified node. Connect an image → vision mode (the image is captioned via Gemma-4 vision = llama.cpp mtmd + auto-downloaded base gemma-4-E2B mmproj). No image → text mode (the prompt, Korean/ English NL or tags, is expanded in-process via llama-cpp-python). Either way the output runs through the same kgen.formatter post-processing (sort / dedup / drop-redundant / count-conflict).

⚠️ Vision mode is NOT an accurate image tagger. It is not a labeling/captioning model that faithfully describes what is in the image. It deliberately exploits the LLM's hallucination: it takes the image only as a loose seed and randomly expands it into a plausible, invented tag set (TIPO-style creative upsampling). The output will include tags that are not actually present in the image and will vary run to run. Use it to generate varied prompt ideas for text-to-image — do not use it when you need exact tags for the given image.

Text mode accepts Korean or English natural language, not just tags:

| Input | Output (example) | | --- | --- | | spring, white blouse, pleated skirt, park | spring, white blouse, pleated skirt, park, outdoors, school uniform, long sleeves, sitting, solo, … | | 빨간 원피스, 밀짚모자, 해변 (Korean) | straw hat, red one-piece swimsuit, sun hat, beach, ocean, swimsuit, outdoors, … |

Inputs:

• prompt: text to convert (Korean/English NL or tags).
• model: a dropdown of the *.gguf files found in ComfyUI/models/gguf/. Pick (auto: download default) to fetch the default model (gemma4-tipo-ko-v2-Q4_K_M.gguf from raspie/gemma4-tipo-ko-gguf) into that folder on first use — a progress bar shows on the node while it downloads (reused afterward, never re-downloaded); or drop your own .gguf into ComfyUI/models/gguf/ and select it (reload the node list to see new files).
• tag_length: target verbosity (very_short … very_long).
• sort: prompt-ordering preset applied by kgen.formatter (tags are categorized special/characters/copyrights/artist/general/quality/meta/rating):
  • danbooru (default): artist, rating, special, characters, copyrights, general, quality, meta.
  • quality_first: quality tags lead.
  • artist_first: artist/character lead.
  • general_only: keep only special + characters + general tags.
  • simple: only special, rating, general.
• temperature: sampling temperature.
• ban_tags: comma-separated tags to strip from the result.
• seed: changes the expansion for variation.
• (optional) image: connect a ComfyUI IMAGE to switch to vision mode (image → tags). Fully automatic — first vision use auto-downloads the prebuilt llama.cpp mtmd binary for your OS/arch (Windows / Linux / macOS · x64 / arm64, no build) and the base gemma-4-E2B mmproj. Nothing to install or configure by hand.
• (optional, vision only) mmproj_path: leave empty — the base mmproj is auto-downloaded. Only set this to point at your own mmproj. gpu_layers: 0 = CPU (default); raise to offload layers to GPU.

Notes:

• Longer / more detailed descriptions yield more and more accurate tags.
• The raw input text (incl. Korean) is not mixed into the tag output.
• Redundant general tags are dropped to save tokens: if a more specific tag with the same head noun exists (e.g. white skirt), the bare skirt is removed. Distinct tags like dress/dress shirt are kept.
• Conflicting person-count tags are collapsed per group: with 1girl present, 4girls / multiple girls are removed (1girl + 1boy is kept — groups are independent).
• On any failure (missing deps, model unavailable, inference error) it falls back to the input text unchanged — it never hard-errors.

Supported environment & requirements

| | | | --- | --- | | OS | Windows / Linux / macOS — anywhere ComfyUI runs and a llama-cpp-python wheel exists (x86-64, Apple Silicon/arm64) | | Node 1 (Seasonal) | Any ComfyUI install. Zero dependencies, no model. | | Node 2 (Gemma TIPO) | CPU-only works (no GPU required); GPU optional for speed | | RAM | min ~8 GB (≈5 GB used for the 4-bit model) · 16 GB+ recommended | | Disk | ~4 GB free (3.2 GB GGUF + deps) | | Speed | CPU inference works but is slow (a single prompt can take minutes); a GPU-accelerated llama-cpp-python build is recommended for speed | | Python | ComfyUI's embedded env; llama-cpp-python >= 0.3.23 (auto-installed/upgraded, one restart prompt) |

Install (Node 2 only)

No manual install is required. The first time you actually run Node 2, its dependencies are auto-installed into the active environment once: tipo-kgen (pure-Python) and llama-cpp-python >= 0.3.23 (from the prebuilt CPU wheel index, so no native build). Older builds cannot load the gemma4 architecture, so an existing too-old llama-cpp-python is upgraded automatically — if that happens mid-session ComfyUI will ask you to restart once (a C-extension can't be hot-reloaded). Node 1 never installs anything.

To pre-install manually instead (optional):

pip install -r requirements.txt

If the prebuilt llama-cpp-python wheel install can't find a wheel for your platform/Python, install it yourself once:

pip install llama-cpp-python --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cpu

References & Credits

| Project | Use | License | | --- | --- | --- | | KGen / TIPO by KohakuBlueleaf | kgen.formatter for tag categorization & prompt sorting | Apache-2.0 | | z-tipo-extension by KohakuBlueleaf | Reference implementation of the TIPO node flow; training-data lineage | Apache-2.0 | | llama.cpp | GGUF inference engine; prebuilt llama-mtmd-cli for vision mode | MIT | | llama-cpp-python | Python bindings for in-process GGUF inference | MIT | | Gemma by Google (google/gemma-4-E2B-it) | Base model family the TIPO GGUF derives from | Gemma Terms of Use | | p-e-w/gemma-4-E2B-it-heretic-ara | Decensored gemma-4-E2B the default Korean TIPO GGUF was fine-tuned from (used so over-censoring doesn't corrupt explicit Danbooru tag output) | Gemma Terms of Use | | unsloth/gemma-4-E2B-it-GGUF by Unsloth | Base gemma-4-E2B mmproj-F16.gguf auto-downloaded for vision mode | Gemma Terms of Use | | raspie/gemma4-tipo-ko-gguf | Default Korean→tags TIPO model auto-downloaded by Node 2 | Gemma Terms of Use |

The TIPO expansion concept ("upsample"/expand short prompts into detailed Danbooru tag sets) originates from KohakuBlueleaf's KGen/TIPO and the z-tipo-extension; the default model's fine-tune training data follows that same lineage.