APNext Feelings

FeelingsPromptNode is the mood dial of dagthomas's "SDXL Auto Prompter" pack. Four dropdowns - loving, creepy, melancholy, and "unnerved" - each hold a long list of emotional words that get appended to your prompt. The idea is to steer the atmosphere of an image: whether a portrait reads as beloved and beaming or dejected and desolate, whether a scene feels eerie or serene.

I'll level with you: of all the builders in this pack, this is the one with the softest touch. An emotion word doesn't render directly - the model has to infer it into lighting, expression, color and posture, and sometimes it just doesn't. "melancholy" nudges toward muted color and a downturned face; it won't restyle a scene the way a film stock or an artist name will. Treat it as a finishing nudge, not a lever, and always back it with something concrete - the emotion word plus "rain on the window, dim light" does far more than the word alone.

How it works

Pure Python, no model, no API key. It joins your prompt with the selected feeling words via the separator (comma default) and returns a string. Each menu leads with None / Random / Multiple Random, so a seed lets it roll a mood for you. As a tag-style injection it fits SDXL-lineage checkpoints best; on LLM-encoded 2026 models the words read as instruction text.

The inputs and outputs that matter

• prompt - your base text; the mood words fold in.
• The four emotion menus: loving (109 - admired, adored, beaming, blessed), creepy (42 - chilling, eerie, dread, disquieting), melancholy (48 - aching, bleak, dejected, desolate) and unnerved (45 - agitated, anxious, bewildered, discombobulated).
• seed - drives Random / Multiple Random.
• attributes (off by default) - adds the pack's extra descriptor words for each entry.

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 pack's requirements.txt pulls a heavy stack - openai, anthropic, google-generativeai, transformers, decord, scipy - because the same repo ships cloud-LLM and video nodes. This node needs none of it; decord is the common install failure, and the builders usually load regardless.

Where people get burned

• Expecting the word to carry the image. An emotion is the least concrete thing you can prompt. If the mood isn't landing, add the physical cues that cause it - lighting, weather, expression, color - rather than stacking more feeling words.
• Contradictory moods. "loving" and "creepy" at once average to a blank stare. Pick one register per image.
• Model era. On SDXL/Illustrious the emotion words behave like tags. On Flux 2 / Z-Image / Qwen, fold the feeling into a plain instruction ("a melancholy portrait in dim window light") instead of appending bare mood tags.
• It's a seasoning node. Reach for Cinematic or Art first to build the look; use Feelings to tilt the final atmosphere.