How to Run Civitai Models Locally — Complete Setup Guide
You downloaded a model from CivitAI. Now what? This guide gets you from downloaded file to generated image in under 30 minutes. Everything runs on your hardware — no cloud, no accounts, no content filters.
What You Need
- GPU: NVIDIA with 4 GB+ VRAM (8 GB recommended)
- RAM: 16 GB minimum, 32 GB recommended
- Disk: 20 GB free + your model files (2-25 GB each)
- OS: Windows 10/11 (recommended), Linux, or macOS (limited GPU support)
- Software: Python 3.10+ and Git (Forge auto-installs Python if needed)
Do not have a CivitAI model yet? Start with How to Download Civitai Models.
Step 1 — Pick Your UI
Three real options in 2026:
- Forge (recommended) — Form-based UI. Familiar layout, 30-50% less VRAM than A1111, supports SDXL and Flux. Best for most users.
- ComfyUI — Node-based workflow editor. Best performance, first to support new models, steeper learning curve. For power users.
- Fooocus — Type a prompt, click generate. Minimal options. For absolute beginners who do not want to tweak settings.
Skip A1111 — Forge is a better version of the same thing with lower VRAM usage.
Step 2 — Install Forge
- Download Forge from GitHub
- Extract the zip to a folder with no spaces in the path (e.g.,
C:\ai\forge\) - Run
webui-user.bat(Windows) orwebui.sh(Linux/Mac) - Wait 10-15 minutes on first run — it downloads Python, PyTorch, and dependencies automatically
When done, your browser opens to http://127.0.0.1:7860 with the Forge UI.
If your antivirus blocks the install, add the folder to your exceptions list. This is a common issue with Python-based tools.
Step 3 — Place Your Model Files
Drop your downloaded .safetensors files into the right folders:
| File Type | Forge Folder |
|---|---|
| Checkpoints | models/Stable-diffusion/ |
| LoRAs | models/Lora/ |
| VAEs | models/VAE/ |
| Embeddings | embeddings/ |
Then click the refresh button in the Forge UI next to the checkpoint dropdown. Your model appears in the list.
Step 4 — Configure for Your GPU
Your VRAM determines which models you can run:
- 4-6 GB VRAM (GTX 1650, RTX 2060 6GB): SD 1.5 models only. Set resolution to 512x512. Use CyberRealistic or Realistic Vision.
- 8 GB VRAM (RTX 3060 8GB, RTX 4060): SDXL models work. Set resolution to 1024x1024. Juggernaut XL, RealVisXL, Pony V6 all run well.
- 12+ GB VRAM (RTX 3060 12GB, RTX 4070 Ti+): Flux models work. Use ComfyUI for best Flux performance.
If you hit "CUDA out of memory" errors, add --medvram to the COMMANDLINE_ARGS line in webui-user.bat and restart.
Step 5 — Generate Your First Image
- Select your model from the checkpoint dropdown
- Type a prompt:
- Realistic:
portrait of a woman, natural lighting, photorealistic, 8k - Anime (Pony/Illustrious):
score_9, score_8_up, score_7_up, 1girl, anime style, detailed
- Realistic:
- Click Generate
Your first image appears in 5-30 seconds depending on your GPU and model size.
Step 6 — Optimize Your Settings
Dial in these settings for better results:
- Sampler: DPM++ 2M Karras (realistic) or Euler a (anime)
- Steps: 20-30 for quality, 10-15 for speed
- CFG Scale: 5-7 for SDXL, 7-9 for SD 1.5
- Resolution: 512x512 for SD 1.5, 1024x1024 for SDXL
- Clip Skip: Set to 2 for Pony and Illustrious models
Verify It Works
You should see a clear, coherent image matching your prompt. If the image is:
- Black: VAE issue. Switch to a BakedVAE model version or download the matching VAE.
- Blurry or low-detail: Steps too low. Increase to 25-30.
- Wrong style: Check that you are using the right prompt format for your model (score tags for Pony, standard prompt for SDXL).
Troubleshooting
- "CUDA out of memory": Model too large for your VRAM. Add
--medvramto launch arguments, or switch to a smaller model. - Black/corrupt images: VAE mismatch. Download the correct VAE or use a BakedVAE checkpoint.
- Very slow (minutes per image): Make sure you are running on GPU, not CPU. Update your NVIDIA drivers. Check that CUDA is detected in the Forge console output.
- LoRA has no effect: Use the correct prompt syntax:
<lora:filename:0.8>— adjust the 0.8 weight up or down. - Pony models look bad: You need quality score tags. Add
score_9, score_8_up, score_7_upat the start of your prompt.
What to Do Next
- Browse models: Civitai Models Download List — curated picks across all architectures
- Realistic models: Best Realistic Civitai Models — Juggernaut, RealVisXL, epiCRealism ranked
- Anime models: Best Anime Models on Civitai — Pony, Illustrious, Animagine compared
Or skip all of this — LocalForge AI ships with Forge pre-installed, popular models included, and zero Python/Git setup. Generate images 5 minutes after download.
