Stable Diffusion 3.5 comes in Large, Large Turbo, and Medium variants on an MMDiT backbone, and the base is heavily safety-tuned. For NSFW you layer community fine-tunes and LoRAs from Civitai and use SD3.5’s specific samplers and CFG. Honestly, mature-focused SDXL forks like Pony, Illustrious, and Lustify are often an easier path than fighting its alignment.
Stable Diffusion 3.5 is Stability AI’s flagship SD line, and technically it is a strong model with a modern MMDiT backbone and good prompt handling. But there is one big thing to understand before you invest time: the base SD 3.5 is heavily safety-tuned. Stability aligned it hard against explicit content, so out of the box it resists NSFW and often mangles anatomy. This guide is therefore about uncensoring it: the variants, setup, the correct settings, the LoRA and fine-tune workflow, and an honest take on when a different model is simply the better tool. Every subject discussed is a fictional adult over 18.
The variants
SD 3.5 ships in three main flavors on the same MMDiT architecture.
| Variant | Size | Character | Best for |
|---|---|---|---|
| Large | 8B | Highest quality, heaviest | Best output, 12GB or more VRAM |
| Large Turbo | 8B | Few-step distilled, fast | Quick iteration, low CFG |
| Medium | ~2.5B | Lighter, faster, lower VRAM | Modest cards, smaller footprint |
Large is the quality flagship. Large Turbo is the distilled speed variant that runs in few steps at low CFG (the same turbo rule you see on other distilled models: do not use high CFG). Medium is the accessible one for smaller GPUs. Pick by your hardware and whether you value quality or speed.
One point that applies across all three: the safety tuning is baked into the base weights of every variant, not just one. So choosing Large over Medium does not get you a more permissive model, it just gets you a bigger one. The alignment is a property of the training, so whichever variant you pick, the uncensoring workflow below is the same. Choose your variant on hardware and quality grounds, and plan to add a fine-tune regardless of which one you land on.
Download and setup
SD 3.5 weights are on Hugging Face under Stability AI (you accept the license on the model page). Download your chosen variant plus the required text encoders and VAE. SD 3.5 uses multiple text encoders, so get the full set the model card specifies.
SD 3.5 runs in both ComfyUI and Forge, so pick your interface.
Folder placement in ComfyUI:
- The SD 3.5 model goes in
ComfyUI/models/checkpoints/ordiffusion_models/. - The text encoders go in
ComfyUI/models/text_encoders/orclip/. - The VAE goes in
ComfyUI/models/vae/.
If you prefer a simpler, more familiar interface than node graphs, Forge works well for SD 3.5. Our Stable Diffusion Forge NSFW setup guide covers that path, and the ComfyUI for NSFW AI complete guide covers the node-based route. For installing checkpoints and LoRAs in general, see how to install NSFW checkpoints.
VRAM
| Variant | Comfortable VRAM |
|---|---|
| Large | 12GB or more (less with GGUF) |
| Large Turbo | 12GB, but fast |
| Medium | 8GB, sometimes less |
Medium is the friendly option for smaller cards. Large and Large Turbo are heavier but still far lighter than the 17B and 20B flagships in this series. GGUF quants exist to stretch VRAM further. For strategies on limited hardware, see the best NSFW checkpoints for low VRAM guide, and AMD owners should read the Stable Diffusion AMD GPU guide.
Settings that matter for SD 3.5
SD 3.5 has its own preferences, and using SDXL habits will disappoint you.
Large and Medium
Use moderate steps (commonly around 20 to 30) and a moderate CFG. Follow the model card’s recommended sampler and scheduler. SD 3.5 responds to a specific sampler pairing, so start with the recommended one before experimenting.
Large Turbo
This is a distilled turbo variant, so the turbo rule applies: low steps (around 4 to 8) and low CFG (around 1 to 2). High CFG breaks distilled models, producing fried, oversaturated output. Do not carry a CFG of 7 over to Large Turbo.
| Variant | Steps | CFG | Sampler |
|---|---|---|---|
| Large | 20 to 30 | moderate | model recommended |
| Medium | 20 to 30 | moderate | model recommended |
| Large Turbo | 4 to 8 | 1 to 2 (low) | model recommended |
For the full theory of how CFG, samplers, and schedulers interact, the NSFW AI CFG and sampler settings guide is the reference to keep open.
The uncensoring workflow
Here is the honest core of using SD 3.5 for adult work: the base model fights you. Stability safety-tuned it against explicit content and against certain anatomy, so raw SD 3.5 tends to refuse, blur, or distort. The fix is community fine-tunes and LoRAs.
- Community fine-tunes. Creators retrain SD 3.5 on adult and anatomy-focused datasets to undo the alignment. A good NSFW fine-tune of SD 3.5 behaves far better than the base weights. Browse these on Civitai; our Civitai NSFW generator guide explains how to find and vet them.
- NSFW LoRAs. On top of a fine-tune (or the base), LoRAs teach specific concepts, styles, or content. Place them in the loras folder and load them with a LoRA node or the Forge LoRA menu.
- Prompting. With a fine-tune loaded, describe your fictional adult subject and scene clearly. The fine-tune does the heavy lifting the base model refuses to do.
If the community fine-tune ecosystem for SD 3.5 is thinner than you hoped, that is a real and common finding, and it leads directly to the honest expectations section.
The baseline safety negative prompt
Always keep a safety negative prompt that blocks disallowed content: child, minor, underage, loli, and shota. Every subject must be a clearly adult, fictional person over 18. This is a hard rule across every model in this series, uncensoring workflow or not. Add standard quality negatives (deformed, extra limbs, blurry) as needed, but the safety terms always stay.
Honest expectations: SDXL forks are often easier
This is the part most guides skip. SD 3.5 is a capable model, but its heavy safety tuning means you spend real effort just to reach a baseline that mature-focused SDXL forks give you immediately. For a lot of NSFW work, those forks are simply the easier and more reliable path.
Consider:
- Pony Diffusion is trained for expressive, explicit-capable output and has a massive ecosystem.
- Illustrious XL and its family excel at anime and illustrated adult content.
- Lustify SDXL is a photoreal adult-focused fork that gets you explicit realism without a fight.
These forks were built by the NSFW community for NSFW work, so they cooperate from the first generation, run on modest SDXL-class hardware, and have deep LoRA support. SD 3.5 can be pushed to good places with the right fine-tune, but if your goal is adult content specifically and you do not have a particular reason to use SD 3.5, starting with an SDXL fork saves hours. Our best Stable Diffusion checkpoints for NSFW roundup lays out the strongest options, and the SDXL vs Pony vs Illustrious comparison helps you choose between them.
The zero-fight option
If you would rather skip the entire censorship battle, a hosted generator sidesteps it completely. A hosted uncensored generator is built for adult output from the start, so there is no alignment to undo, no fine-tune hunting, and no setup. It is a sensible option when your priority is results over tinkering, or when your hardware is limited. If you enjoy the local control and want to tune every setting yourself, running SD 3.5 with a good fine-tune is still perfectly viable; it is a matter of how much of your time you want the base model’s alignment to cost.
You can also test prompt ideas free with no install on the free NSFW AI generator here.
A realistic prompt walkthrough
To make this concrete, here is how a prompt evolves on SD 3.5 with a good NSFW fine-tune loaded. Start with the subject and frame it as a clearly adult fictional character: describe age-appropriate features, build, and expression first. Then layer wardrobe, then the environment, then lighting, then the camera and style. A structured prompt might move from the character, to a specific outfit, to a bedroom or studio setting, to soft window light, to a photographic 50mm look.
The reason this ordering matters on SD 3.5 specifically is that the MMDiT backbone weights early tokens heavily. Front-load the elements that define the image and leave stylistic garnish for the end. If you bury the main subject behind a long stack of quality tags, the model gives you a technically clean image of the wrong thing. With a fine-tune doing the uncensoring, your job is simply to be clear and orderly.
When a generation comes out close but not right, change one thing and reroll rather than rewriting the whole prompt. Adjust the lighting clause, or swap the camera language, or nudge the LoRA weight up or down by a small amount. Iterative single-variable changes teach you how this particular fine-tune responds, which is more valuable than chasing a perfect prompt in one shot.
Combining fine-tunes, LoRAs, and upscaling
A mature SD 3.5 workflow usually stacks three things: a NSFW fine-tune as the base, one or two LoRAs for specific concepts or styles, and an upscale pass for the final export. Keep LoRA weights moderate. Two LoRAs at full strength often clash and produce artifacts, so dial each back and test. If a LoRA has trigger words, include them, and if it does not, its effect applies just by being loaded.
For the finishing upscale, generate at the model’s native resolution first, pick your keepers, then run a hires pass only on those. Our NSFW AI hires fix complete guide explains the denoise-and-upscale loop, and the NSFW AI upscaler guide compares dedicated upscale models. On Large Turbo, remember to keep CFG low on the upscale pass too, or you will reintroduce the fried-contrast problem the low-CFG base run avoided.
If you want to build your own SD 3.5 LoRA rather than rely on what is available, the how to train a NSFW LoRA guide covers the full process. Training on SD 3.5 is viable, though the community has concentrated more effort on SDXL forks, so datasets and examples are easier to find there.
Strengths of SD 3.5
- Solid MMDiT architecture with good prompt handling once uncensored.
- Three variants covering quality, speed, and low VRAM.
- Runs in both ComfyUI and Forge, so you can use a familiar interface.
- A real fine-tune of it can produce excellent, coherent adult output.
Limitations to expect
- Heavy safety tuning. The defining issue. The base model resists NSFW and anatomy, so you must fine-tune or LoRA your way past it.
- Thinner adult ecosystem. SD 3.5’s NSFW fine-tune and LoRA library is smaller than the enormous SDXL fork ecosystem, so you have fewer ready-made options.
- Effort versus payoff. For pure adult work, the setup cost often exceeds what an SDXL fork demands for equal or better results.
- Turbo discipline. Large Turbo needs low CFG; treat it like any distilled model.
If generations come out blurry, distorted, or refuse to render your prompt, that is usually the base alignment showing, and the fix is a proper NSFW fine-tune, not more steps. For other issues, see the Stable Diffusion not working fix and the general NSFW AI troubleshooting guide.
When SD 3.5 is actually the right choice
Given all the honesty above, when should you pick SD 3.5 over an SDXL fork? A few genuine cases. If you already like the SD 3.5 look and a fine-tune exists that matches your style, there is no reason to switch. If you want the newer MMDiT architecture’s prompt handling and are willing to source a good fine-tune, SD 3.5 can produce cleaner adherence than older SD 1.5 era models. And if you specifically want the Medium variant’s light footprint on a smaller card while staying on the SD 3.5 line, it is a reasonable modern base.
What SD 3.5 is not is a shortcut. If your only goal is fast, reliable adult output and you have no attachment to this specific model, the SDXL forks win on ecosystem depth and cooperation. Treat SD 3.5 as a capable model you can uncensor with effort, not as the path of least resistance. For a direct sense of how the older and newer SD generations compare on adult work, the SD 1.5 vs SDXL comparison is a useful frame, since it shows how much the base training and community focus shape real-world NSFW results more than raw architecture does.
First-run checklist
- Pick a variant: Large for quality, Large Turbo for speed, Medium for low VRAM.
- Download the variant plus all text encoders and the VAE from Hugging Face, accepting the license.
- Set up ComfyUI or Forge and place the files in the correct folders.
- Download an SD 3.5 NSFW fine-tune and any LoRAs from Civitai.
- Use the correct settings: moderate steps and CFG for Large and Medium, low steps and low CFG for Large Turbo.
- Load the fine-tune, write a clear prompt describing your fictional adult subject, and keep the safety negative prompt in place.
- If the base fights you too much, switch to an SDXL fork like Pony, Illustrious, or Lustify.
- If you want to skip the whole fight, use a hosted uncensored generator instead.
SD 3.5 is a good model wearing a strong safety collar. With the right fine-tune and settings you can absolutely produce quality adult content on it, but be honest about the effort. For many creators the mature-focused SDXL forks reach the same destination faster, and a hosted uncensored generator skips the setup entirely. Choose the path that matches how much time you want to spend versus how much control you want to keep.
Frequently asked questions
Is SD 3.5 censored?
Yes, the base SD 3.5 is heavily safety-tuned by Stability AI against explicit content and certain anatomy, so out of the box it resists NSFW, blurs, or distorts. You uncensor it with community fine-tunes and LoRAs from Civitai. Be aware its adult ecosystem is thinner than SDXL’s, which is why many creators prefer mature-focused SDXL forks.
How much VRAM does SD 3.5 need?
It depends on the variant. Large and Large Turbo (both 8B) want around 12GB or more for comfort, less with GGUF quants. Medium at roughly 2.5B runs on about 8GB and sometimes less. All three are lighter than the 17B and 20B flagship models, so SD 3.5 is relatively accessible hardware-wise.
Which SD 3.5 variant is best for NSFW?
Large gives the best quality if you have the VRAM and pair it with a good NSFW fine-tune. Large Turbo is faster but needs low CFG discipline. Medium suits smaller cards. The variant matters less than the fine-tune you load on top, since the base alignment is the real obstacle regardless of variant.
What CFG and steps should I use for SD 3.5?
For Large and Medium, use moderate steps around 20 to 30 with a moderate CFG and the model’s recommended sampler. For Large Turbo, it is distilled, so use low steps (4 to 8) and low CFG (1 to 2). High CFG breaks turbo variants with fried, oversaturated output, so never carry an SDXL value of 7 over.
Is it easier to use Pony or Illustrious instead of SD 3.5?
Often yes. Mature-focused SDXL forks like Pony, Illustrious, and Lustify were built by the NSFW community, so they cooperate from the first generation without fighting alignment, run on modest hardware, and have deep LoRA support. For pure adult work with no specific reason to use SD 3.5, an SDXL fork usually saves hours.
Is SD 3.5 free?
Yes, SD 3.5 weights are free to download from Hugging Face after you accept Stability AI’s license on the model page, and you run it locally at the cost of hardware and power. Community fine-tunes and LoRAs on Civitai are also generally free. A hosted uncensored generator is the paid convenience alternative.
How do I get NSFW output from SD 3.5?
Load a community NSFW fine-tune of SD 3.5 that retrains away the safety alignment, then optionally stack LoRAs for specific concepts. Browse and vet these on Civitai, place them in the right folders, and prompt clearly. The fine-tune does the work the base model refuses. Always keep the safety negative prompt in place.
Does SD 3.5 run in Forge?
Yes. SD 3.5 runs in both ComfyUI and Forge, so you can use a familiar image-style interface instead of node graphs if you prefer. Make sure your Forge build is current enough to support the SD 3.5 architecture and its multiple text encoders, then load the checkpoint and any NSFW fine-tune the same way you would an SDXL model.
