ComfyUI NSFW Setup Guide
Put SDXL checkpoints in ComfyUI/models/checkpoints/ and LoRAs in ComfyUI/models/loras/, then restart ComfyUI so Load Checkpoint and Load LoRA see new files. Install ComfyUI Manager next — it is the fastest way to close “missing node” gaps when you import a workflow JSON from CivitAI or Discord.
The Models
1. Single-file SDXL checkpoint
Top PickFastest first image — keep files under models/checkpoints/.
Architecture: SDXL .safetensors · VRAM: 8 GB+ typical @ 1024² FP16 · Best for: Load Checkpoint → CLIP encode → KSampler → VAE decode
View on CivitAI →2. ComfyUI Manager
Use Manager before you hand-edit custom_nodes/ for common packs.
Architecture: custom_nodes / registry · VRAM: N/A · Best for: Installing missing nodes from imported workflows
View on CivitAI →3. Flux / GGUF split loaders
Different folders + GGUF nodes — not a drop-in SDXL swap.
Architecture: Quant UNET + split CLIP/VAE · VRAM: 4–12 GB (quant-dependent) · Best for: Flux-class models when SDXL is not the target
View on CivitAI →Why This Matters
ComfyUI fails in boring ways: files in the wrong folder, workflows built with nodes you never installed, or VRAM plans that assume a 4090 when you have 8 GB. This guide is the decision-stage checklist: paths → Manager → first errors → VRAM reality — NSFW is just the same loaders with different weights.
Install Path
Pick one lane and stay in it:
| Lane | What you get | Tradeoff |
|---|---|---|
| Git + venv | Reproducible Python, easy updates | You own pip conflicts |
| Portable / embedded Python | Fewer global-Python surprises | Heavier folder, still need Manager deps on some builds |
| Desktop app | Manager often on by default | Less transparent file layout |
If you want the stack without living inside pip every weekend, LocalForge AI is one managed local option — same ComfyUI idea, less env drift.
Folder Map
| Asset | Path | Notes |
|---|---|---|
| Checkpoints | ComfyUI/models/checkpoints/ |
SDXL full .safetensors |
| LoRA | ComfyUI/models/loras/ |
Chain after base model |
| VAE | ComfyUI/models/vae/ |
Swap if colors drift |
| ControlNet | ComfyUI/models/controlnet/ |
Match preprocessor nodes |
| Embeddings | ComfyUI/embeddings/ (build-dependent) |
Textual inversion |
Share a disk with A1111 / Forge? Copy extra_model_paths.yaml.example → extra_model_paths.yaml in the ComfyUI root and point at shared folders — no duplicate terabytes.
ComfyUI Manager
- Desktop: Manager may already be enabled — open the UI and search node packs before you git-clone random repos.
- Portable Windows: install
manager_requirements.txtwith the same Python that launches ComfyUI, then run with--enable-managerper current docs. - Missing nodes: import the workflow JSON → use Install Missing when the prompt appears — then restart.
First-Run Errors
| Symptom | Fix |
|---|---|
| Checkpoint missing in dropdown | File not under models/checkpoints/ or ComfyUI not restarted |
| Red “missing node” graph | Install pack via Manager or add the custom node repo under custom_nodes/ |
| OOM at 1024² | Drop resolution, batch 1, disable live preview, close parallel browser tabs on same GPU |
| Wrong colors / gray cast | Wrong VAE — point VAELoader at the VAE that matches the checkpoint family |
| LoRA does nothing | LoRA after model load; strength > 0; update ComfyUI if loaders changed |
Low VRAM / Flux-Class Note
SDXL NSFW in this cluster is mostly single-file checkpoints. Flux / GGUF is a different graph: quantized UNETs land under models/diffusion_models/, CLIP/VAE split loaders, and ComfyUI-GGUF-style nodes. If you are on 8 GB and trying Flux, plan for quant + fp8 text encoders — do not expect SDXL defaults to carry over.
Mid-page CTA: Want help keeping deps aligned while you experiment? LocalForge AI packages local workflows so you spend less time fixing Python and more time iterating.
What to Do Next
- Pick weights: Best NSFW Models for ComfyUI — ranked CivitAI checkpoints with ComfyUI loading notes.
- Pick a UI: ComfyUI vs A1111 for NSFW — graphs vs WebUI before you double-buy models.
- Tool context: ComfyUI for NSFW — routing when you outgrow basics.