Local AI Girlfriend: Why Is This So Hard to Set Up?

An AI girlfriend is three separate things, each with its own tool stack:

1. The AI model (the brain). Something that generates text — the model that decides what she says. Tools: KoboldCPP, Oobabooga (text-generation-webui), Ollama, LM Studio. These are model runners — they load and serve the AI model. By themselves they have no personality; they're just the text-generation engine.

2. The chat character interface (the face). Something that gives her a persona, memory, chat UI, character cards. Tools: SillyTavern (dominant), RisuAI, Agnai. These are chat frontends — they give you an interface to talk to a character. By themselves they do nothing; they need a model to connect to.

3. The image generation stack (the body). Something that generates photos of her. Tools: ComfyUI, Automatic1111, Fooocus running Stable Diffusion / FLUX. These are image generators — separate from text generation entirely. For "same woman every photo," you also need per-persona LoRA training via Kohya (yet another tool).

Nobody warned you these are three separate stacks that need to be stitched together. The tutorials assume you already know this, so when you install KoboldCPP expecting an AI girlfriend and get a chat interface with no character, you think you did something wrong. You didn't; KoboldCPP just isn't that product — it's one of three pieces you need.

KoboldCPP. Runs a text AI on your computer. Loads a model file (GGUF format), serves text generation on a local port. Just the brain, no face. Technical: a C++ llama.cpp wrapper, single-executable download, works on CPU and GPU.

Ollama. Also runs a text AI, with a friendlier CLI for downloading and managing models. Just the brain, no face. Technical: Go-written wrapper around llama.cpp with model-pulling via a package-manager-like interface.

LM Studio. Also runs a text AI, with a desktop GUI for managing models. Just the brain, has a built-in basic chat UI that's minimal (no character, no memory, no persona). Technical: Electron app wrapping similar inference libraries.

Oobabooga (text-generation-webui). Also runs a text AI with a web UI. Has somewhat more features than LM Studio's built-in chat (character support exists but is less polished than SillyTavern). Technical: Python/Gradio app with pluggable backends.

SillyTavern. Chat frontend. Gives you character cards, persistent memory, group chat, rich chat UI. Does NOT run the AI model itself — it points at whichever model runner you have serving on a local port, or at a hosted API. Just the face, needs a brain.

ComfyUI / Automatic1111 / Fooocus. Image generators. Separate from everything above. For "images of your specific character," also need a per-character LoRA trained via Kohya.

The glue work. Even with all tools installed and running, you need to configure SillyTavern to connect to your model runner, build or import a character card for your persona, train a LoRA for her face, build a workflow that calls image generation when you want it. Add another 10-20 hours of integration work on top of the setup of each individual tool.

`illegal instruction (core dumped)` or `no AVX2 support`. Your CPU is too old to run the AI inference library at full speed. Common on older laptops. Fix: find a version of the tool with AVX fallback, or run on GPU instead of CPU, or get newer hardware.

`model file not found` or `can't find model`. You installed the tool but didn't download the actual AI model file (usually a .gguf or .safetensors file in the gigabyte range). Fix: find a model on Hugging Face, download it to the tool's models directory, point the tool at it.

`CUDA out of memory`. The model you're trying to run is bigger than your GPU VRAM. Fix: use a smaller model, use a more aggressively quantized version (Q4_K_M instead of Q8_0, for example), or run on CPU (much slower).

`ModuleNotFoundError` in Python-based tools. The Python environment doesn't have required dependencies installed. Fix: usually means the venv wasn't set up correctly; delete the venv folder and relaunch the tool to rebuild.

`failed to load model` without more specifics. Often a model-format incompatibility (you downloaded a model in a format the tool doesn't support). Fix: verify the model file is in the expected format (GGUF for KoboldCPP/Ollama/LM Studio, safetensors for most ComfyUI workflows).

`SillyTavern connects but no response`. SillyTavern is pointing at a model-server address that isn't serving a model. Fix: verify your model runner (KoboldCPP, Oobabooga, etc.) is actually running and listening on the port SillyTavern expects.

All of these are standard first-time setup errors. None of them are catastrophic once you know what they mean. The problem is beginners don't know what they mean, and the error messages assume you do.

Hosted AI companion apps bundle brain + face + image gen + voice + memory into one subscription. The whole reason the category exists is that nobody was DIY-stitching this in large numbers because it's too much work for most users.

Sloane Premium at $19.99/mo includes: chat with 80+ hand-curated personas or custom-built ones (model + character all in one), persistent memory of who you are, per-persona photos (image gen + LoRA already trained), voice notes from the persona, short video clips. All one product. Sign up in 15 minutes.

Other hosted companion apps in the category:

Candy AI. Larger roster of anime and realistic characters. Image-first strength. See Candy AI review.

Character.AI. Massive roster of user-generated characters. The 2023 content policy shift hurt the product for many users. See will Character.AI allow NSFW? for current context.

Replika. Original AI companion product. The 2023 ERP removal fractured the user base; current state is more wellness-focused. See Replika alternatives after 2023 update for context.

Nomi, Kindroid, DreamGF, Muah AI, Pipsqueak. Smaller companion apps with specific angles. Varying strengths; worth comparing if the top three don't fit.

All of these productize the chat + memory + persona + images stack that local setups make you assemble yourself. Price is similar across the category ($10-30/mo for most). Setup is minutes vs. 20-60 hours.

Three situations where local is the correct answer despite the setup burden.

Privacy is a hard requirement. With local, your prompts and conversations stay on your machine; nothing leaves. Hosted companion apps have the vendor seeing your prompts, which matters if "my data can't leave my machine" is a non-negotiable constraint for your use case.

You plan to use it 10,000+ messages/month. Hosted monthly subscriptions include generous but not infinite usage. At high volumes (chat-constantly-every-day usage patterns), local avoids the ongoing cost ceiling. The breakeven depends on specific usage patterns; at very high volumes local wins on long-term economics.

You enjoy the tooling itself. If setting up the stack is part of what you got into this for — if the process is a hobby, not a cost — the setup time doesn't count as a drawback. Build the stack, iterate on it, enjoy the control. For users in this category, local wins because the "cost" axis of time investment is actually a benefit.

Outside of those three situations, hosted companion apps productize what you're trying to assemble. Reasonable people make different calls based on which situation they're in; both answers are defensible.

For "I want it to work tonight," any of the top three (Sloane, Candy AI, Character.AI) is a reasonable landing spot. Try a couple for a month before committing; flat monthly pricing makes experimentation cheap.