🎨 Auto Prompt RedNote

Every so often a node ships that's clearly one person's very specific obsession, and Auto Prompt RedNote is that node. It generates prompts in the aesthetic of Xiaohongshu (RedNote) - the Chinese social app where everything is ethereal, sparkly, softly lit, and aggressively safe-for-work - and it does it with a mood slider and a two-model switch. It's the most opinionated node in the comfyui-auto-anime-prompts pack, and easily the most interesting.

You reach for it when you want that specific "aesthetic girl in a dreamy scene" look and you don't want to hand-tune the enforcement tags yourself. The README calls it "highly aesthetic, social-media style generations with managed presets," which is accurate right down to the word managed - this node enforces its look.

How it works

Two completely different generation paths, selected by target_model:

• Illustrious (Tags) - the tag path. Stacks quality tags, a RedNote style suffix, your character, a random action/background/camera, a mood tag, and a character-enforcer suffix that pins things like (solo:1.5), (perfect cute face:1.4), (flat chest:1.2), rating_safe. If the random action involves sitting/hugging/lying, it also appends (pretty white lace safety shorts:1.3) to keep the platform-safety look.
• Flux/Qwen (Natural) - the sentence path. A clean_tag() pass strips weight syntax (:1.2), parens, underscores, and 1girl out of the tag soup, then reassembles everything into natural sentences glued by connectors: "A high-quality anime illustration of Miku, a girl with blue hair, twintails, aqua eyes. She is currently eating strawberry crepe. The scene takes place in a sunny beach. Her expression is …" This is the fix for "tag soup" - the exact failure mode you hit when you feed Illustrious-style tags to an LLM-encoded model like Flux.

The mood_level slider (0.0–1.0, default 0.5) is the signature feature. It picks a mood tag bucket: below 0.2 it's demure and smiling, at 0.5 it's a stoned, dissociated hollow gaze, at 1.0 it's (stubborn:1.5), (pouting:1.4), (grumpy:1.4), (angry:1.2). The author tuned that curve like a person who has thought about the emotional range of anime girls more than is entirely healthy.

Inputs worth knowing

• preset - the standard eight plus a new RedNote default that layers in the enforced aesthetic and, on the negative side, a heavy safety blocklist ((large breasts:1.5), nsfw, nude, loli, …). In Flux mode the negative comes out empty, which is correct - Qwen/Flux-class models do better with a blank negative than with booru garbage.
• enable_style_lock - locks the style file to the same index as the character, so character #5 always gets style #5. Off by default, which means styles are rolled randomly per prompt.
• mode - sequential / random character selection.
• seed - defaults to -1 here (true random, time-seeded), unlike the other nodes' default 0. Set it if you want a reproducible run.

Outputs: prompt (list), negative, character_name (list), and mood_tags (list) - the last one is handy for showing exactly which mood bucket the slider picked, and for wiring into an easy-showAnything debug node.

Install & caveats

Manager search "Auto Anime Prompt", or clone it into custom_nodes:

cd ComfyUI/custom_nodes
git clone https://github.com/jluo-github/comfyui-auto-anime-prompts

No extra deps, no model downloads. The list-output gotcha from the batch/combiner nodes applies here too: prompt is a list, so a plain CLIP Text Encode sees only slot 0.

The honest caveat: this is the most opinionated node in the pack, and opinionated means it can fight you. The RedNote enforcers will override your own composition ideas - (flat chest:1.2) doesn't care what you think about proportions. If you want the aesthetic without the aesthetic police, set preset to something other than RedNote and use only the mood slider. Given how niche the whole thing is, that's my move: borrow the mood system, skip the enforcement.