LoRAs, Checkpoints, VAEs & Embeddings Explained — Beginner Download & Install Guide
Confused by Stable Diffusion model files? Here's what checkpoints, LoRAs, VAEs, and embeddings actually are, where to download them, and exactly where to put them. No jargon, no assumptions.
The Jargon Problem
You decided to run Stable Diffusion locally. You picked an interface. Now someone tells you to "download a checkpoint, grab a LoRA, swap your VAE, and add some embeddings." What does any of that mean?
This guide explains every model type in plain English, tells you exactly where to download them, and shows you where each file goes on your machine.
Shortcut: LocalForge AI ships with the best checkpoints, LoRAs, and VAEs pre-installed and pre-configured. If you'd rather skip the manual setup, it's ready out of the box.
Checkpoints — The Core Model
What It Is
A checkpoint is the main AI model — the thing that actually generates images. It's a large file (2–7 GB) containing billions of learned parameters. Think of it as the "brain" of the operation.
Types You'll See
- SD 1.5 checkpoints (~2 GB): the classic. Huge ecosystem, thousands of fine-tunes. Works on 4 GB VRAM.
- SDXL checkpoints (~6.5 GB): higher quality, better anatomy. Needs 6+ GB VRAM.
- Flux checkpoints (~12 GB): newest generation. Best prompt adherence. Needs 8+ GB VRAM.
Where to Download
- CivitAI — largest collection, community ratings, NSFW models available
- Hugging Face — official model repos, research models
Where to Put It
Drop the .safetensors file into:
models/Stable-diffusion/
Then select it from the checkpoint dropdown in your UI. That's it.
LoRAs — Style & Subject Add-Ons
What It Is
A LoRA (Low-Rank Adaptation) is a small add-on file (10–200 MB) that modifies a checkpoint's output without replacing it. LoRAs teach the model new styles, characters, poses, or concepts.
Examples
- A LoRA trained on a specific art style (e.g., "watercolor illustration")
- A LoRA trained on a specific character or face
- A LoRA for better hands, better lighting, or specific poses
- An "add detail" LoRA that sharpens output across any checkpoint
Where to Put It
Drop the .safetensors file into:
models/Lora/
Then reference it in your prompt: <lora:filename:0.8> — the number (0.0–1.0) controls strength.
Pro Tip
You can stack multiple LoRAs in one prompt. Start with low weights (0.5–0.7) and increase if the effect is too subtle. Too high and the image will look "fried."
VAEs — Color & Clarity
What It Is
A VAE (Variational Autoencoder) handles the final decoding step — converting the AI's internal representation into an actual image. A good VAE produces sharper colors, less washed-out images, and better fine details.
Do You Need One?
Most checkpoints include a VAE baked in. But if your images look washed out or have dull colors, swapping to a dedicated VAE file fixes it instantly. The most popular is vae-ft-mse-840000 for SD 1.5 models.
Where to Put It
Drop the .safetensors or .pt file into:
models/VAE/
Then select it in Settings → Stable Diffusion → SD VAE.
Embeddings (Textual Inversions) — Prompt Shortcuts
What It Is
An embedding is a tiny file (10–80 KB) that teaches the model a new concept or keyword. You type the embedding name in your prompt and it injects a learned concept into the generation.
Common Uses
- Negative embeddings: "EasyNegative" and "bad-hands-5" — put them in your negative prompt to automatically avoid common defects
- Style embeddings: trigger a specific artistic style with a single keyword
- Concept embeddings: teach the model objects or characters that aren't in the base training data
Where to Put It
Drop the .safetensors or .pt file into:
embeddings/
Then use the filename as a keyword in your prompt. Example: adding EasyNegative to your negative prompt.
Quick Reference Table
| Type | Size | Folder | What It Does |
|---|---|---|---|
| Checkpoint | 2–12 GB | models/Stable-diffusion/ | The core model — determines overall style & capability |
| LoRA | 10–200 MB | models/Lora/ | Add-on for styles, characters, concepts |
| VAE | 300–800 MB | models/VAE/ | Improves color accuracy & sharpness |
| Embedding | 10–80 KB | embeddings/ | Prompt shortcut for concepts or quality fixes |
The Starter Stack (What to Download First)
If you're setting up manually, here's the minimum recommended stack:
- Checkpoint: Juggernaut XL (photorealism) or DreamShaper XL (versatile) from CivitAI
- LoRA: "Add More Details" LoRA — works with any checkpoint
- VAE: sdxl_vae.safetensors (for SDXL models) or vae-ft-mse-840000 (for SD 1.5)
- Embeddings: EasyNegative + bad-hands-5 for negative prompts
Total download: roughly 8–10 GB. Total setup time if you already have an interface installed: 30–60 minutes (mostly download time).
Or: LocalForge AI includes an optimized starter stack out of the box — checkpoints, LoRAs, VAE, and embeddings pre-configured and ready to generate in under 10 minutes.
Common Mistakes
- Using an SD 1.5 LoRA with an SDXL checkpoint: won't work. LoRAs must match the base model architecture.
- LoRA weight too high: anything above 1.0 usually produces artifacts. Start at 0.6–0.8.
- Wrong folder: if the UI doesn't show your model, it's in the wrong directory. Double-check the paths above.
- Forgetting the VAE: if images look washed out or have a grey/purple tint, you need to set a VAE.
- Downloading .ckpt instead of .safetensors: always prefer
.safetensors— it's the safe, modern format..ckptfiles can contain executable code.