You hit Queue Prompt, the progress bar finishes, and the preview is a flat black square. Or a blurred mess. Or an image that looks nothing like what you asked for. If you searched for a ComfyUI NSFW filter disable switch, you have probably already noticed that there is no tidy toggle in the settings menu. The truth is more interesting: the filter is usually a node you added yourself, a model baked with its own restrictions, or a plain technical fault that only looks like censorship.
This article walks through all three cases in plain language. You will see where the black images come from, how to bypass or remove a safety checker node, what to check when the output is still broken, which uncensored model families are worth loading, and when a hosted option is simply faster than fighting your own graph.
💡 Quick answer: ComfyUI core does not ship a global NSFW filter. If your images are blacked out, look for a Safety Checker node in the workflow, then check VAE, precision and model mismatches before blaming censorship.
What the ComfyUI Filter Really Is

ComfyUI Core Has No Filter
A clean install of ComfyUI is a node graph runner. It loads a checkpoint, encodes your prompt, samples a latent and decodes it with a VAE. Nothing in that chain looks at the finished picture and decides whether it is acceptable. That is why you will not find a "disable NSFW" checkbox anywhere in the interface or the launch arguments.
What you can find are three separate layers that people lump together:
| Layer | Where it lives | Can you switch it off? |
|---|
| Safety Checker custom node | A node pack you or a downloaded workflow added | Yes, bypass or delete it |
| Model-level restrictions | Training data or tensors inside the checkpoint | Only by using a different or fine-tuned model |
| Cloud API moderation | The provider behind an API node | No, it is controlled by the provider |
Most "my ComfyUI is censored" threads turn out to be the first row. A shared workflow includes a checker node, the user never notices it, and every borderline image gets replaced.
Where Black Images Come From
The best known culprit is the CLIP-based Safety Checker node pack. It is essentially the old Stable Diffusion 1.x safety filter packed into one node, and when it flags an image it replaces it with black. This is destructive: if you only use its IMAGE output, the original pixels are gone.
A second node pack, YetAnotherSafetyChecker, behaves a little differently. It outputs black unless the input falls below the threshold you choose, so a threshold set too low blacks out perfectly harmless pictures. Several maintained forks of the original also exist, which is worth knowing because a workflow that "worked yesterday" may pull a different fork today.
And then there are the impostors. A pure black image with no error message very often comes from NaN values in the latent before the VAE decode, which has nothing to do with content at all. We cover those fixes in a later section.
Disable the Safety Checker Node

You have three practical ways to switch the checker off. Pick the lightest one that solves your problem.
Bypass With Ctrl+B
Click the Safety Checker node on the canvas and press Ctrl+B. The node turns purple and ComfyUI passes the input straight through to the next node as if the checker were not there. This is the safest option because you can undo it with the same shortcut and the rest of the graph stays intact.
After bypassing, check the wiring. Your Save Image or Preview Image node must take its input from the VAE Decode output, or from whatever sits right before the checker, not from the checker output. If the checker was the only thing feeding your save node, re-link it manually.
Delete the Node Pack
If the node keeps coming back when you load shared workflows, remove the cause instead of the symptom:
- Open the ComfyUI Manager and find the installed safety checker pack.
- Uninstall it, or delete its folder inside
custom_nodes.
- Restart ComfyUI fully, not just the browser tab.
- Reload your workflow. Missing nodes will be highlighted in red, so delete those from the canvas and reconnect the save node.
💡 Tip: Keep a clean "base" workflow saved separately. Downloaded workflows often bundle extra node packs you never asked for, and a base graph lets you compare outputs in seconds.
Tune the Threshold
Sometimes you want to keep a checker for a shared machine but stop the false positives. In that case raise the tolerance instead of removing the node. With YetAnotherSafetyChecker the threshold is the first value to change, since it decides what counts as acceptable. With the CLIP-based node, confirm that the sensitivity is set the way you intend and that the node is actually wired to the image input you think it is.

Not Working? Fix These First
You bypassed the node, restarted, and the image is still black. Good news: this almost always points to a technical fault, and technical faults have boring, reliable fixes.
Black Image Without a Checker
Work through this short list in order:
- Check the console. Warnings about NaN, overflow or "invalid value encountered" point at precision, not moderation.
- Disable every custom node you do not need. A conflict between packs can produce the same blank output as a filter.
- Test with a default workflow. Load a plain text to image graph with one checkpoint. If that works, your original graph is the problem.
- Lower resolution and batch size. Memory pressure sometimes yields garbage frames.
Wrong VAE or Precision
The VAE decode step is the most common hidden cause of black frames. Here is what the community has reported working:
| Symptom | Likely cause | Fix |
|---|
| Pure black with XL-class checkpoints | NaN latent before decode | Restart ComfyUI with --fp32-vae |
| Same symptom, want speed | fp16 VAE overflow | Load an fp16-fix VAE through a Load VAE node |
| Black with Flux checkpoints | Wrong or empty VAE file | Load the Flux VAE (ae.safetensors) and re-download if it is zero bytes |
| Black or green on GTX 16-series cards | fp16 defect on that hardware | Launch with --force-fp32 |
| New 2026 base models output black | Model-specific bug reported by users | Check the model's issue tracker for the current workaround |
The fp16-fix VAE helps XL-class checkpoints but does nothing for Flux Dev style models, which need their own VAE. Running the VAE on CPU also works, but it is slow enough that most people only use it to confirm the diagnosis.
LoRA and Base Mismatch
Mixing families is the other classic. A LoRA trained for XL-class models will not work on Flux, and loading it with the wrong loader can give you noise, distortion or a broken graph. Check the base model listed on every LoRA page before you wire it in. The same rule applies to ControlNets, IP-Adapters and VAEs: one family per graph.
💡 Rule of thumb: If the image is black, suspect the VAE. If it is garbled, suspect a LoRA or sampler mismatch. If it is blurred or replaced after generation, suspect a checker node.
Uncensored Models Worth Loading

Once the filter is out of the way, the model itself becomes the limit. Base checkpoints released by large labs are usually trained with adult material filtered out or with built-in restrictions, so they tend to refuse, soften or distort certain subjects even with no checker node present.
Flux Family Options
Flux Dev is loved for realistic skin, lighting and hands, but the base release is conservative. The community answer has been fine-tunes: for example, a Flux.1-dev fine-tune that swaps out the safety tensor and adds extra adult training. Quality varies from release to release, so test a handful of prompts before committing to a download.
If you like building custom looks, LoRA-based workflows are the lighter route. Flux Dev LoRA and Flux 2 Klein 9B Base LoRA show how a small adapter can shift style and subject matter without retraining the whole model.
XL Checkpoints From the Community
Most uncensored checkpoints live on Civitai, and models marked with the red NSFW badge are generally trained without safety filters. The catch is compatibility: an XL-class checkpoint needs XL-class LoRAs, VAEs and ControlNets. If you are curious how a fast XL-class model behaves, SDXL Lightning 4Step is a handy reference for how few steps a modern pipeline actually needs.
For photographic realism with a gentler learning curve, Realistic Vision v5.1 remains a popular pick for lifelike portraits, and Z Image Turbo is a quick option when you want sharp results in seconds.
What to Check Before Downloading
- Source and license. Read the model card. Some licenses restrict commercial use.
- File format. Prefer
.safetensors over pickle-based files.
- Base family. Match it with your LoRAs, VAE and workflow.
- Recent comments. Users usually report black outputs and broken samplers within days.
Be realistic about the research here: much of the written advice on "best uncensored checkpoints" comes from promotional blogs, and no neutral ranking exists. Treat every list, including this one, as a starting point and test on your own prompts.

Cloud Nodes and Their Own Rules
ComfyUI can also call hosted models through API nodes. Those requests leave your machine, and the provider applies its own moderation on its own servers. No node setting, bypass shortcut or launch flag changes that. If an API node refuses a prompt or returns a blank result, the fix is a different model or provider, not a different graph.
This is the point where many people give up on the local route for adult-themed work. Local generation gives you control but costs time: downloads, VRAM limits, VAE hunts, node conflicts. A hosted model that is already permissive removes that whole list.

How to Get Similar Results on PicassoIA
ComfyUI itself is not a PicassoIA model, so there is nothing to launch here. But if your goal is attractive, tasteful, adult-themed imagery without maintaining a node graph, these hosted models are the faster path. Seedream 4.5 comes first because it handles glamour, swimwear and editorial portraits with strong realism and very little friction.
Step by Step With Seedream 4.5
- Open the Seedream 4.5 page and choose the text to image mode.
- Write the subject first, then the setting, then the light. Example: "woman in a black silk slip dress on a linen sofa, morning window light from the left, 85mm lens, soft film grain."
- Choose a 16:9 ratio for wide editorial frames or 3:4 for portraits.
- Pick the highest resolution offered. The model is built for 4K output, so fine skin texture and fabric detail survive cropping.
- Generate, then change one variable at a time: lens, light direction or wardrobe. This is the same discipline you use when debugging a ComfyUI graph.
💡 Prompt tip: Name real photographic details such as lens focal length, light direction and film stock. Vague words like "beautiful" or "sexy" give flat results.

Edit and Iterate With Image Editor Pro
When an image is nearly right but the outfit, background or pose is off, PicassoIA Image Editor Pro is the next stop. It offers unlimited generations, which matters when you iterate twenty times on one idea. Describe the change in plain language, keep the rest of the image as a reference, and repeat until the result matches your picture in your head. No nodes, no VAE files, no restarts.
Other Models in Order

Safe Habits for Adult Content
Switching off a filter removes a safety net, so replace it with your own habits:
- Adults only, fictional subjects. Never generate images of real people without their consent, and never depict anyone who looks underage. This is a hard line.
- Keep it tasteful if you plan to publish. Platforms, ad networks and payment processors have rules. Suggestive, glamour-style images travel much further than explicit ones.
- Store files privately. Generated outputs should live in a folder you control, not a shared drive or synced workspace.
- Label AI images where required. Many platforms ask for disclosure, and some regions are adding legal requirements.
- Do not share workflows blindly. A workflow downloaded from a stranger can contain custom nodes that run code on your machine.
If you share your own graphs, strip out anything personal: file paths, model names tied to private LoRAs, and embedded metadata you would rather keep to yourself.
Make Your Own Images Today

Here is the short version of everything above. If ComfyUI blacks out your images, bypass the Safety Checker node with Ctrl+B, delete the node pack if it keeps returning, and rule out the VAE, precision and LoRA mismatch before assuming censorship. If the model itself resists, move to a fine-tune built for it, or skip the plumbing entirely.
Local graphs are great when you enjoy tinkering. When you just want a stunning photo of a person in a sunlit loft or on a quiet beach, a hosted model is quicker. Open Seedream 4.5, write one detailed prompt, and see how far a single sentence can go. Then take the best result into PicassoIA Image Editor Pro and refine it as many times as you like.
Want to compare more options? Browse the full catalog at picassoia.com/en/all-models and try a few prompts of your own. The best way to find your favorite look is to generate, tweak and repeat.