Stop writing wildcard files by hand
Wildcard Studio is a free Windows desktop app that builds your wildcard file for you. Paste any text, drop in a reference image, or grab tags from a booru post — a local AI model pulls out every visual detail and sorts it into categories.
Why should you use an application to write a wildcard file
It's tedious. A file worth using needs hundreds of entries across dozens of categories, typed one line at a time.
The formatting fights you. One unbalanced brace doesn't throw an error — it rides straight into your prompt as a stray character. Bad indentation silently drops a whole category. You find out when a roll comes back as garbage.
You can't remember what's in it. Past a few hundred entries you're adding duplicates or avoiding adding at all.
The categories drift.
twintailsgoes under Hair.Style today and Hair.Type next month, so neither one ever fills up properly.It stays flat. By hand you write whole looks as single entries like
long messy blonde twintails— one line that fires as one thing. Split into length, texture, colour and style, those same four words become dozens of combinations you never typed.It's capped by your own recall. You write the details you can think of in the moment and stop. The file mirrors your imagination instead of widening it.
It's never finished. Every new idea means going back in and hand-filing it, forever.
Application Highlights
It knows what you already have
This is the core of the app: every extraction is compared against your existing wildcard file before anything is proposed. Keywords you already have are recognized and greyed out — only genuinely new material is offered. Your collection grows with every source you feed it, without ever collecting duplicates.
That's what makes it different from pasting text into ChatGPT and asking for keywords: a chatbot doesn't know what's in your file. Wildcard Studio reads your file every run, so it stays useful on the hundredth source, not just the first.
Five tools in one application
Extract:
Paste any text (copied prompt, clothing description, anything). The model breaks it into atomic keywords, each filed to a category like Hair.Color or Clothing.Legwear. Review each proposal, edit the word or its category, and add what you want in one click.
Caption:
Load an image or provide an image URL and a local vision model describes it in three formats: booru tags, prompt phrases, and natural language. One click sends the result to Extract to become wildcard entries.
Booru
Search for booru tags to add or paste a post URL, get the full tag list as a ready-made prompt, and feed it straight into your wildcard file.
Editor
A tree editor for the file itself. No model needed. Allows you to manually add/delete a catagory / subcategory to your wildcard file
Preview
Use your wildcard file and see real prompts come out. Validates the syntax and flags the mistakes wildcards otherwise swallow silently, like unbalanced braces.
Preview uses Dynamic Prompts syntax — the same {a|b|c} and __wildcard__ forms understood by Stable Diffusion's Dynamic Prompts extension and ComfyUI. Anything that rolls correctly here rolls the same way there.
That means you can build and test a prompt entirely inside Wildcard Studio, then paste it straight into your generator. And unlike those processors, Preview tells you when you've made a mistake: unbalanced braces and wildcards pointing at categories that don't exist both get flagged instead of quietly corrupting your prompt.
Requirements
Windows
LM Studio or Ollama running a local model — vision model needed for Caption
Editor and Preview need nothing installed at all
Which model to use
Which model to use
Not sure? Do this. Install LM Studio, download Mistral-Small-3.2-24B-Instruct-2506 from unsloth (or the largest model in the list below that fits your graphics card), set its context length to 32768, and switch on the Local Server. That's the whole setup. Everything below is for tuning — you don't need it to get started.
4 GB VRAM — Qwen2.5-3B-Instruct (Q4_K_M). Works, but expect broader categories and more hand-correcting. See the note below.
8 GB — Qwen2.5-7B-Instruct or Llama-3.1-8B-Instruct (Q4_K_M)
12 GB — Mistral-Nemo-12B-Instruct or Qwen2.5-14B-Instruct (Q4_K_M)
16 GB — Mistral-Small-3.2-24B-Instruct-2506 (Q3_K_XL, ~11 GB) — best all-rounder
24 GB+ — the same at Q5 or Q6
Mistral-Small-3.2-24B also sees images, so from 16 GB up one model covers both Extract and Caption. Get it from unsloth — their _XL quants beat plain ones of the same size and include the vision mmproj file Caption needs.
Small or no graphics card? Extract and Caption are one-shot jobs — you click once and wait — so running slowly is perfectly usable here, unlike a chatbot. On 4 GB, turn LM Studio's GPU offload slider down and run a 7B or 12B model partly on system RAM instead of picking a 3B model: much slower, noticeably better results. With no usable GPU at all, run entirely on CPU — a 7B model will finish an extraction in a minute or two on a normal desktop.
You may also need to drop context to 8192 or 16384 on a small card, since the context itself uses VRAM. Paste a paragraph at a time rather than a whole page to compensate.
Editor and Preview need no model at all, so building and rolling your wildcard file works on any machine.
Working with adult material
Standard instruct models are trained to refuse or soften explicit content. In Extract that shows up as a run that returns far fewer keywords than the text contained, or nothing at all — it isn't a bug, the model just declined. In Caption you'll get a vague description that skips whatever it didn't want to name.
Try the standard model first. Mistral-Small-3.2-24B is fairly permissive and usually handles adult source text without complaint. Most people don't need anything else.
If it refuses, use an abliterated build. These have the refusal behaviour stripped out. huihui-ai publishes them on Hugging Face for most popular models.
Don't reach for one by default. Abliteration degrades instruction-following, and Extract leans on that heavily to produce clean categories. You'll trade sharper filing for fewer refusals, so only switch if you're actually being blocked.
For Caption, use a purpose-built captioning fine-tune. There are vision models on Hugging Face trained specifically to describe adult images in booru-tag and natural-language form, and they're far better at it than a general VLM. Just don't use one for Extract — captioning fine-tunes drift into prose and ignore the format Extract needs.
Avoid thinking models for Extract (the reasoning step silently drops keywords) and captioning fine-tunes for Extract (they ignore the required format).
Set context length to 32768 before your first run — it's the most common cause of a failed extraction.
Before you download
The app is unsigned, so Windows shows a "Windows protected your PC" warning the first time. Click More info → Run anyway. Some antivirus tools flag PyInstaller-built apps as suspicious — a known false positive.
Full instructions are in HELP.html, included in the download.
Feedback welcome
This is v1.0 and I want to make it better. If you try it, tell me which model you ran and how the categories came out — that's the part most worth tuning.
MIT licensed.