How to Install NSFW Checkpoints in 2026 (A1111, Forge, ComfyUI)

To install NSFW checkpoints in 2026, download the safetensors file from Civitai, then drop it in your front end’s model folder: models/Stable-diffusion for Automatic1111 and Forge, or models/checkpoints for ComfyUI. Click refresh, select the model, and generate. VAE files go in a separate VAE folder. The whole process takes about two minutes.

We install checkpoints constantly across multiple rigs, and the process is genuinely simple once you know the right folder for each tool. This guide walks through downloading from Civitai, the exact folder paths for Automatic1111, Forge, and ComfyUI, where VAE files go, how to refresh the model list, and how to fix the two classic failures: a model that will not load and a model that produces black images. If you do not want to install anything at all, you can generate adult images right now through our browser tool, which needs no setup, no GPU, and no downloads.

Step 1: Download from Civitai

Most NSFW checkpoints live on Civitai. On any model page you will see a green Download button. A few rules we always follow:

• Always download the safetensors format, never the older ckpt format. Safetensors cannot execute hidden code, so it is the safe choice.
• Note the base model listed on the page (SD1.5, SDXL, Pony, or Illustrious). This determines which resolutions and LoRAs are compatible.
• If the page offers a pruned or fp16 variant, grab it. It is smaller and identical in quality for generation.
• Check whether the page says the VAE is baked in. If it is, you do not need a separate VAE file. Most modern checkpoints bake it in.

Civitai requires you to be logged in and to have enabled mature content in your account settings to see and download NSFW models. That is a one-time toggle in your profile settings.

Step 2: Find your front end’s model folder

This is the step that trips up almost everyone, because the three major tools use different folder names. Here is the definitive map.

Front end	Checkpoint folder	VAE folder	Refresh method
Automatic1111	models/Stable-diffusion	models/VAE	Recycle icon by checkpoint dropdown
Forge	models/Stable-diffusion	models/VAE	Recycle icon by checkpoint dropdown
ComfyUI	models/checkpoints	models/vae	Refresh button or reload tab

The key gotcha: Automatic1111 and Forge use a folder literally named Stable-diffusion, while ComfyUI uses checkpoints. People who switch tools often drop the file in the wrong place and then wonder why the model never appears. There is no error message for this; the model is simply invisible.

Full paths in context

Inside your installation directory the full paths look like this:

Automatic1111:
  stable-diffusion-webui/models/Stable-diffusion/your-model.safetensors
  stable-diffusion-webui/models/VAE/your-vae.safetensors

Forge:
  stable-diffusion-webui-forge/models/Stable-diffusion/your-model.safetensors
  stable-diffusion-webui-forge/models/VAE/your-vae.safetensors

ComfyUI:
  ComfyUI/models/checkpoints/your-model.safetensors
  ComfyUI/models/vae/your-vae.safetensors

Move or copy the downloaded .safetensors file into the correct checkpoint folder. You do not need to unzip anything; a checkpoint is a single file.

Step 3: Place the VAE (only if needed)

A VAE, or variational autoencoder, decodes the model’s latent output into the final image. Most modern checkpoints have the VAE baked in, so you can skip this step. You only need a separate VAE if:

• The Civitai page explicitly says no VAE is included, or
• Your images come out washed out, gray, or with weird color, which signals a missing or wrong VAE.

When you do need one, the common files are sdxl_vae.safetensors for SDXL-based models and vae-ft-mse-840000 for SD1.5-based models. Drop the VAE in the VAE folder from the table above, then select it in your front end’s settings. In A1111 and Forge, that is Settings, then VAE, where you can set it to Automatic or pick a specific file. In ComfyUI, you load it with a separate Load VAE node and wire it into your decode step.

Step 4: Refresh the model list

You do not need to restart the whole application after adding a model. Each tool has a quick refresh:

• Automatic1111 and Forge: click the small recycle or refresh icon next to the checkpoint dropdown at the top left. The new model appears in the list. Select it and wait a few seconds for it to load into VRAM.
• ComfyUI: click the Refresh button in the menu, or simply reload the browser tab. Then pick the model in your Load Checkpoint node.

If the model still does not show up after refreshing, it is in the wrong folder. Go back to Step 2 and double check the path. This is the cause roughly nine times out of ten.

Step 5: Apply the right settings and generate

Load the model, then match your settings to its base type, because the bases want different settings:

• SD1.5 models: generate at 512×512 or 512×768, Clip Skip 1 or 2 depending on the model, CFG 5 to 7.
• SDXL models: generate at 1024×1024 or 832×1216, Clip Skip 1 (off), CFG 4 to 7.
• Pony-based models: generate at SDXL resolutions, Clip Skip 2, CFG 5 to 7, and use the score tag ladder in your prompt.
• Illustrious models: SDXL resolutions, Clip Skip 2, and tag-based prompting.

Generating an SDXL or Pony model at SD1.5 resolutions like 512×512 is a top cause of broken anatomy and duplicated limbs, so always match resolution to the base. If you would rather skip the settings learning curve entirely, our hosted generator is pre-tuned and lets you generate without touching any of this.

Paths on Windows, Linux, and macOS

The folder structure is identical across operating systems; only the path separators and root location differ. On Windows your install often sits somewhere like C:\Users\you\stable-diffusion-webui\models\Stable-diffusion. On Linux it is typically ~/stable-diffusion-webui/models/Stable-diffusion. On macOS it is the same as Linux, usually under your home folder. ComfyUI follows the same pattern with its models/checkpoints directory. Whatever the OS, the rule does not change: the folder name is what matters, not the drive or root. If you use a file manager, navigate into the tool’s install directory, then into models, and you will see the subfolders waiting for you.

A handy trick for people who run more than one front end: you can point all of them at a single shared models folder so you do not store duplicate multi-gigabyte files. In ComfyUI, edit the extra_model_paths.yaml file to reference your A1111 models directory. In A1111 and Forge, use the --ckpt-dir launch argument to point at an external folder. This keeps one library on disk that every tool reads from, which is a major space saver once your collection grows past a few models.

Downloading large models reliably

NSFW SDXL and Pony checkpoints are large, often 6 to 7GB, and a dropped connection mid-download produces a corrupt file that fails to load with a confusing error. A few habits prevent this:

• Use a download manager for anything over a couple of gigabytes. It can resume an interrupted transfer instead of forcing a restart.
• Verify the file size against the number shown on the Civitai page once the download finishes. A short file is an incomplete file.
• Check the hash if Civitai lists one. Matching the SHA256 confirms the file is intact and untampered.
• Avoid renaming mid-download. Let the transfer finish, then rename the completed file to something descriptive.

If you have a Civitai account, the site also offers an API and various community download helpers that handle resumes automatically, which is worth setting up if you pull models frequently.

Troubleshooting: model will not load

If a checkpoint fails to load or throws an error, work through these in order:

• Wrong folder: the most common cause. Confirm the file is in models/Stable-diffusion (A1111/Forge) or models/checkpoints (ComfyUI).
• Incomplete download: a half-downloaded file will fail. Check the file size against the Civitai listing; SD1.5 models are roughly 2GB, SDXL and Pony models roughly 6 to 7GB. Re-download if it is short.
• Out of VRAM: an SDXL model on a small card can fail to load. Launch with the --medvram flag in A1111 or Forge, which streams parts of the model to save VRAM.
• ckpt vs safetensors mismatch: some old extensions choke on one format. Stick with safetensors.
• Stale model list: you forgot to refresh. Click the recycle icon or reload the tab.

Troubleshooting: black or NaN images

A black image, or a console error mentioning NaN, almost always means a VAE precision problem on certain NVIDIA GPUs, not a corrupt model. The fixes, in order of preference:

• Add --no-half-vae to your launch arguments in A1111 or Forge. This forces full-precision VAE decoding and resolves black images on most cards with negligible speed cost.
• Try --no-half if the above does not work, though it is slower.
• Load an external VAE in fp32 if the baked VAE is the problem.
• In ComfyUI, confirm your VAE Decode node is wired to a valid VAE, either the checkpoint’s built-in VAE via the VAELoader or a separate one.

For 1000 series and some 1600 series GPUs, you may also need --precision full --no-half, though this is rare on current drivers.

Troubleshooting: washed out or gray images

If images are not black but look gray, flat, or desaturated, the VAE is missing or mismatched. Add the correct VAE for the base model: an SDXL VAE for SDXL and Pony models, an SD1.5 VAE for SD1.5 models. Set it to Automatic in A1111 settings, or wire the correct Load VAE node in ComfyUI. This usually restores proper contrast and color immediately.

Organizing a growing model folder

Once you have more than a handful of checkpoints, a little organization saves a lot of time:

• Use subfolders. A1111, Forge, and ComfyUI all support subfolders inside the checkpoint directory, so you can group by base type, such as SDXL, Pony, and SD15. They show up as categories in the dropdown.
• Keep names descriptive. Rename files to include the base, like cyberrealistic-pony-v18.safetensors, so you remember which settings each one wants.
• Prune duplicates. Fp16 and pruned variants are identical in quality to full versions for generation, so delete the heavier copies to save disk space.
• Track VAE needs. Keep a short note of which of your models need an external VAE, so you are not chasing washed out output later.

Once your library grows, our checkpoint roundup is a useful reference for matching each model to the settings and base type it wants, so you can keep a lean, purposeful folder instead of hoarding fifty merges.

A note on NSFW content and local front ends

Unlike many hosted services, the open-source front ends covered here do not apply a content filter to your generations. Automatic1111, Forge, and ComfyUI generate exactly what the loaded checkpoint and your prompt produce, with no built-in safety classifier blocking adult output. That is precisely why local installation is the standard route for NSFW work: the checkpoint itself determines what it can render, and an NSFW-trained checkpoint like the ones on Civitai will produce adult content without any extra configuration. There is no hidden toggle to flip inside the UI; install an NSFW checkpoint, prompt it, and it works.

The only filter you will encounter is on Civitai itself, where you must enable mature content in your account settings to browse and download these models. Once the file is on your machine and in the right folder, the front end treats it like any other checkpoint. Keep your library private and follow your local laws regarding adult content, but technically the install process is identical to installing any SFW model.

First generation checklist

Before you hit generate on a freshly installed model, run through this quick list to avoid the most common first-run disappointments:

• File is in the correct folder for your front end, and you clicked refresh.
• Resolution matches the base type, native SDXL sizes for SDXL and Pony, 512-based sizes for SD1.5.
• Clip Skip is set correctly, 2 for Pony and Illustrious, off for standard SDXL.
• Prompt format matches the base, score tags for Pony, natural language plus tags otherwise.
• A VAE is present, either baked in or selected, so output is not gray.
• The no-half-vae flag is set if you have previously seen black images on your GPU.

Tick those six boxes and your first generation will almost always look the way the model’s showcase images do, instead of a broken or washed out mess that sends you back to forums.

Our verdict

Installing NSFW checkpoints is a two-minute job once you know the one rule that matters: the checkpoint folder is models/Stable-diffusion for Automatic1111 and Forge, but models/checkpoints for ComfyUI. Get that right, refresh the list, match your settings to the base type, and you are generating. The two failure modes, models that will not load and black images, almost always trace back to a wrong folder or a VAE precision flag, both of which are quick fixes. And if you ever want to skip the local setup entirely, our browser generator is always there as a no-install option to try a model’s look before you commit disk space and VRAM to it.

Frequently asked questions

Where exactly do NSFW checkpoints go in Automatic1111?

In Automatic1111, checkpoint files go in the models/Stable-diffusion folder inside your webui installation directory. The full path is stable-diffusion-webui/models/Stable-diffusion/your-model.safetensors. After copying the file there, click the small recycle icon next to the checkpoint dropdown at the top left to refresh the list, then select your model. Forge uses the exact same folder name, which trips up people who assume it differs.

Why is ComfyUI’s checkpoint folder different from A1111?

ComfyUI uses models/checkpoints while Automatic1111 and Forge use models/Stable-diffusion. They are just different naming conventions chosen by different developers. This mismatch is the single most common installation mistake when people switch tools: they drop the file in the folder name they remember from their old front end, and the model silently never appears because there is no error for a file in the wrong place.

Do I always need to install a separate VAE?

No. Most modern NSFW checkpoints in 2026 have the VAE baked into the model file, so you can skip VAE installation entirely. You only need a separate VAE if the Civitai page explicitly says none is included, or if your images come out washed out, gray, or oddly colored, which signals a missing VAE. When needed, use an SDXL VAE for SDXL and Pony models, and an SD1.5 VAE for SD1.5 models.

Why are my generated images completely black?

Black images almost always mean a VAE precision problem on certain NVIDIA GPUs, not a corrupt download. The fix is to add the no-half-vae argument to your launch command in Automatic1111 or Forge, which forces full-precision VAE decoding. This resolves black output on most cards with negligible speed cost. If that fails, try no-half, or load an external fp32 VAE. In ComfyUI, check your VAE Decode node wiring.

My new checkpoint does not appear in the dropdown. Why?

Nine times out of ten the file is in the wrong folder. Confirm it is in models/Stable-diffusion for Automatic1111 or Forge, or models/checkpoints for ComfyUI. The second most common cause is forgetting to refresh: click the recycle icon by the checkpoint dropdown, or reload the ComfyUI tab. A third possibility is an incomplete download, so check the file size matches the Civitai listing.

Safetensors or ckpt, which should I download?

Always choose safetensors. The older ckpt format can contain executable code that runs when the model loads, which is a genuine security risk with files from the open internet. Safetensors is a safe container that cannot execute anything, and it loads just as fast with identical quality. Every reputable NSFW checkpoint on Civitai offers a safetensors version, so there is no reason to ever download ckpt.

What settings should I use after installing a model?

Match settings to the base model. SD1.5 models generate at 512×512 with CFG 5 to 7. SDXL models want 1024×1024 with Clip Skip off. Pony-based models need SDXL resolutions, Clip Skip 2, and the score tag ladder in the prompt. Generating an SDXL or Pony model at 512×512 causes duplicated limbs and broken anatomy, so resolution matching is the most important setting to get right.

Can I organize checkpoints into subfolders?

Yes. Automatic1111, Forge, and ComfyUI all support subfolders inside the checkpoint directory, and they appear as categories in the model dropdown. Grouping by base type, such as SDXL, Pony, and SD15 folders, makes a large library manageable and reminds you which settings each model wants. Pair this with descriptive filenames that include the version and base, and a growing model collection stays easy to navigate.