APNext Art

Where the Artist node gives you who, this one gives you how - the medium, the technique, the surface. It's a stack of dropdowns for art types, painting styles, sculpture materials, color palettes, patterns, and technique descriptors, and it appends whatever you pick onto your prompt. If your problem is that everything comes out looking like the same glossy default render, this is a fast way to shove it toward gouache, or a byzantine mosaic, or a 1950s pulp sci-fi cover. It ships in dagthomas's "SDXL Auto Prompter" pack alongside the Artist, Cinematic and Character builders.

The lists are genuinely large. The types menu alone is 192 entries deep - abstract expressionism, airbrush painting, altered books, on and on - and styles, palette, patterns and techniques each run to dozens more. That's the appeal and the trap: it's a great way to discover a medium you'd never have typed, and an easy way to pile on ten conflicting descriptors and get an averaged nothing.

How it works

No model and no API - it concatenates your prompt with the selected terms using the separator (comma by default) and hands back a string. Every menu starts with None / Random / Multiple Random, so the node doubles as a randomizer: set a seed, flip a menu to Random, and each run rolls a different medium. This is a CLIP-era, tag-style approach, which means it's a natural fit for SDXL-lineage checkpoints; on LLM-encoded 2026 models the terms still read as text but land as instruction fragments rather than weighted tags.

The inputs and outputs that matter

• prompt - your base text; the picks get folded in.
• The medium menus that carry the most weight for a beginner: types (192 - the broad art-form list), styles (63 - parenthetical looks like (baroque painting), (anime), (byzantine mosaic)), and techniques (97 - brushwork, etching, tool and effect descriptors). sculpture, palette, patterns and painting are there when you want material, color-scheme, surface or media-combo specificity.
• seed - powers the Random / Multiple Random picks.
• attributes (off by default) - folds in the extra descriptor words the pack attaches to entries.

Outputs are prompt and random, both strings, into a CLIP Text Encode.

How to install it

ComfyUI Manager: search comfyui_dagthomas, install, restart. Or manually:

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

then restart. The requirements.txt is bigger than this node needs - it drags in transformers, decord, scipy, openai, anthropic and friends because the same pack ships cloud-LLM and video nodes. decord is the usual install headache; if it fails, the pure-Python builders like this one generally still load.

Where people get burned

• Descriptor soup. Turning on five menus at once is the fastest way to make an image less coherent. Each term dilutes the others. Change one thing at a time and keep the seed fixed.
• Overlap with styles. Several menus can describe the same axis (a type, a style, and a technique all fighting over "how painterly"). Let them contradict and you get averaged mud - this is exactly the "contradictory prompt" failure mode. Pick the one that best names what you want.
• Model era. On SDXL/Illustrious this behaves like the tag list it is. On Flux 2 / Z-Image / Qwen, keep it short and concrete - a named medium works, a pile of vague style words doesn't.