AI Enhance#
AI enhancement rewrites your prompt into the house style of the model you picked before the image is generated. It is off by default, and it is controlled by three separate switches.
The three controls#
There is no "enhancement strength" or "enhancement level" picker. What exists is:
| Control | Where | Type | Default |
|---|---|---|---|
enhance |
parameter on /imagine, /turbo, /compose, /render |
True / False / unset | unset (falls back to your setting) |
raw |
parameter on /imagine and /turbo, also a saved setting |
True / False | False |
ai_features |
/settings → Features → AI |
4-state cycle | Disabled |
Leaving enhance unset means "use my ai_features setting". Setting it explicitly overrides that setting for one job. raw: True beats everything and disables enhancement outright.
Parameter descriptions in Discord#
| Command | What the enhance option says |
|---|---|
/imagine |
Auto-enhance prompt with AI |
/turbo |
Auto-enhance prompt with AI |
/compose |
Enhance your description into a full caption with AI (uses AI features setting by default) |
/render |
Rewrite your prompt into MiniMax H3's timeline format with AI (default: your AI features setting) |
The ai_features setting#
Open /settings, go to Features, and click AI to cycle through four states:
| Icon | State | Character matching | LLM enhancement |
|---|---|---|---|
| 🚫 | Disabled | off | off |
| 🎭 | Match Only | on | off |
| ✨ | Enhance Only | off | on |
| 🌟 | Both Enabled | on | on |
The cycle order is Disabled → Match Only → Enhance Only → Both Enabled → Disabled. Disabled is the default, so out of the box nothing is enhanced and nothing is matched.
Note
enhance: True on a command turns on LLM enhancement only. It does not turn on character matching -- that is controlled solely by ai_features. If you want /imagine miku to resolve to Hatsune Miku, you need Match Only or Both Enabled.
Which models enhance by default#
The enhance flag is interpreted differently per model family.
| Model family | Behaviour |
|---|---|
| SDXL / Illustrious / NoobAI / Pony | strict opt-in |
| Lumina2 / Z-Image / Neta-Lumina | strict opt-in |
| Mugen | strict opt-in |
| Flux 2 | strict opt-in |
| Flux Dev | strict opt-in |
| Krea 2 | strict opt-in |
| Qwen Image | strict opt-in |
| ERNIE Image | default ON -- enhances unless you pass enhance: False |
| Mage-Flow | default ON -- enhances unless you pass enhance: False |
| Ideogram 4 | always on -- force-enabled regardless of your ai_features setting |
Ideogram 4 needs its enhancement
Ideogram 4's enhancement is what produces the structured caption the model actually reads. Passing enhance: False or raw: True disables it and produces noticeably worse output. See Prompting Tips.
Which AI does the rewriting#
| Order | Provider | Runs where | Notes |
|---|---|---|---|
| 1 | Google Gemini (gemini-2.5-flash) |
cloud | primary enhancer for every command |
| 2 | Local LLM | on the bot's CPU | fallback whenever Gemini can't or won't |
| -- | TIPO | inside ComfyUI | tag expander used on tag-based anime checkpoints instead of an LLM |
Gemini is always tried first. The local model takes over when:
- no Gemini key is configured on the bot,
- Gemini has been marked degraded by the health monitor,
- Gemini times out or errors on both of its attempts, or
- Gemini refuses the prompt on content-policy grounds -- the local model is deliberately used to pick those up.
If both fail, the job runs with your original prompt.
Gemini runs on the maintainer's own API key, so whether enhancement is available at all depends on that key still having credit. The local model is the fallback precisely because it does not.
There is no GPT / OpenAI enhancement path
OpenAI is used only for OpenAI image models in /edit. It never touches your prompt text.
The local model is also the only provider used for LoRA trigger placement, which weaves a selected LoRA's trigger words into a sensible position in your prompt. That call is capped at 3 seconds and is skipped silently if it runs long.
What the enhancer is given#
Your prompt is not sent alone. The enhancer also receives:
| Context block | Contents |
|---|---|
| Model instructions | The hand-written rule set for the family you picked (tags for Illustrious, prose for Flux, and so on) |
| Active LoRAs | Each selected LoRA's nickname, description and trigger words, with an instruction not to write the trigger words itself |
| Detected characters | Any matched characters, with their canonical appearance and required tag format. With two or more characters it also gets multi-character formatting guidance |
| Possible art styles | If your wording names a known artist, franchise or art movement, a description of that style is offered as a suggestion the model may ignore. At most 10 candidates |
The enhancer is explicitly told never to emit quality tags (masterpiece, score_9, 8k) or LoRA triggers, because the bot adds those itself. See how your prompt is assembled.
Timing#
Enhancement runs in the queue worker, not when you submit the command, so it never delays your "job queued" confirmation. It does add time to the job itself: Gemini allows up to 30 seconds per attempt and two attempts, so a bad network moment can cost up to a minute before the local model takes over. The local model on CPU is slower than Gemini.
Re-rolls, edits and upgrades inherit the parent job's enhanced prompt and are not re-enhanced, so your prompt stays stable across a 🔄 re-roll.
When enhancement fails#
Failures are completely silent
If enhancement fails, the job runs with your original prompt. There is no error message, no warning embed, and no note on the result card. The only sign is the absence of the enhanced marker.
On top of that, a provider that fails 3 times in a row is marked degraded and skipped entirely on new jobs, while a background probe retries it on a backoff that grows from 10 seconds up to 10 minutes. One success restores it.
So if your prompts suddenly stop being enhanced, or the enhanced results start reading in a noticeably different voice, that is the fallback chain at work -- not a change you made.
How to tell enhancement ran#
| Where | What you see |
|---|---|
| Queue / status line | the word enhanced in the parameter summary |
| Details panel | prompt: with your original, plus enhanced: with the rewritten version |
| Details panel (matching only) | matched characters: listing the characters that were recognised |
/compose |
an enhanced lyrics: line |
The Details panel also records which enhancer ran -- Gemini, TIPO, or the local model by name.
Match Sensitivity#
The Match Sensitivity value in /settings controls character matching, not enhancement. Default 0.85, range 0.0--1.0.
The tooltip is wrong
The settings menu describes it as "Lower = stricter matching, higher = more matches". The actual behaviour is the opposite: the value is a minimum similarity score, so higher is stricter and produces fewer matches. Lower it if names are not being recognised; raise it if the bot matches characters you did not mean.