The best upscaler models for NSFW AI art in 2026 are 4x-UltraSharp and 4x_foolhardy_Remacri for photoreal skin, 4x_NMKD-Siax and RealESRGAN_x4plus as reliable all-rounders, and 4x-AnimeSharp or RealESRGAN_x4plus_anime_6B for anime. Download them from OpenModelDB or Hugging Face and drop them in models/ESRGAN or models/upscale_models. These are deterministic and add no new detail. Keep all subjects adult, fictional, and AI-generated.
When people say “upscaler” they sometimes mean the app (Upscayl, ComfyUI) and sometimes mean the model file the app loads. This guide is about the model files: the .pth and .safetensors ESRGAN-architecture weights you download and feed to your upscaler. Picking the right one is the difference between crisp, natural skin and a smeared or over-sharpened mess. If you want the apps instead, see best NSFW AI upscalers.
One thing to understand up front: these models are deterministic. They enlarge an image by a fixed factor (usually 4x) and sharpen the existing detail, but they invent nothing new. A soft source stays soft, just bigger. That makes them fast, repeatable, and completely uncensored, since they run locally and only process what is already in the image. To actually add detail you pair them with a diffusion pass, covered in how to add detail to NSFW AI images.
Where to get the models
Two trusted sources:
- OpenModelDB (openmodeldb.info) is the community catalog. It lists hundreds of ESRGAN-architecture models with sample images, scale factors, and recommended use cases. This is the best place to browse and compare.
- Hugging Face hosts many of the same models, especially the official Real-ESRGAN releases. Good for direct, stable download links.
Download the .pth or .safetensors file, not a zip of training data. File sizes are small, typically 60MB to 70MB for a 4x model.
Where to put the models
The folder depends on your app.
# Automatic1111 / Forge
stable-diffusion-webui/models/ESRGAN/
# ComfyUI
ComfyUI/models/upscale_models/
# Upscayl (custom models)
Upscayl settings -> custom models folder you choose
# chaiNNer
Load from any folder via the model load nodeAfter dropping a file in, refresh or restart the app and the model appears in the upscaler dropdown or node. If it does not show up, you are pointed at the wrong folder, which is the single most common setup mistake. See the ComfyUI complete guide and Forge setup guide for paths.
The model comparison table
| Model | Best for | Character | Notes |
|---|---|---|---|
| 4x-UltraSharp | Photoreal, general | Sharp, clean, slightly crisp | The most popular default for realistic AI art |
| 4x_foolhardy_Remacri | Photoreal skin, portraits | Soft, natural, low halo | Excellent on skin, avoids the plastic look |
| 4x_NMKD-Siax_200k | General photoreal | Balanced, detailed | Great all-rounder, handles varied content |
| RealESRGAN_x4plus | General | Neutral, reliable | The standard baseline, ships with many tools |
| RealESRGAN_x4plus_anime_6B | Anime, flat color | Clean lines, flat areas | Official anime model, light and fast |
| 4x-AnimeSharp | Anime, illustration | Crisp lines, vivid | Sharper than the 6B for line art |
| SwinIR (4x) | Photoreal, restoration | Smooth, transformer-based | Slower, strong on noise reduction |
| 4x_NMKD-Superscale | General, textures | Detail-heavy | Good for fabric and surface texture |
Photoreal models in detail
4x-UltraSharp
The default reach-for choice for realistic AI art. It produces a clean, crisp result with good edge definition. It can be a touch aggressive on already-sharp sources, which shows as slight halos, so pair it with a diffusion pass at 0.3 denoise for the best skin. For most photoreal upscales this is where to start.
4x_foolhardy_Remacri
The skin specialist. Remacri is softer and more natural than UltraSharp, with minimal haloing, which makes it the favorite for portraits and any image where skin texture matters. If UltraSharp looks too harsh on a face, switch to Remacri. It is the antidote to the plastic look.
4x_NMKD-Siax_200k
A superb all-rounder from the NMKD model family. It balances sharpness and detail without the aggressiveness of UltraSharp, and it handles mixed content (skin, fabric, hair, background) gracefully. A safe default if you upscale a wide variety of images.
RealESRGAN_x4plus
The baseline that ships with countless tools. Neutral and reliable, it rarely produces a bad result, but it is also rarely the best for any single category. Use it when you want predictable, no-surprises output.
SwinIR
A transformer-based architecture rather than classic ESRGAN. It is smoother and particularly good at noise reduction and gentle restoration, at the cost of speed. Reach for it when your source has compression noise or grain you want cleaned up.
Anime and illustration models
RealESRGAN_x4plus_anime_6B
The official lightweight anime model. It keeps flat color areas flat and lines clean, exactly what 2D art needs. It is small and fast. This is the default for upscaling output from Pony Diffusion or Illustrious.
4x-AnimeSharp
Sharper and more vivid than the 6B, better when you want crisp, punchy line art. It can slightly over-emphasize edges on softer illustrations, so test both on your style. For clean cel-shaded work it is excellent.
The key rule: never use a photoreal model on anime or an anime model on a photo. A photoreal model adds skin-like texture to flat anime color, creating noise and grain where there should be smooth fill. An anime model on a photo flattens skin into a cartoon. Match the family to the content every time.
Deterministic vs diffusion: the crucial distinction
This trips up a lot of people. These ESRGAN models do not add detail. They are pixel enlargers with smart sharpening. If your 512px source has no visible pores, no upscaler model will create them; it will just make the smooth skin bigger.
To genuinely add detail you need a diffusion upscale: run the enlarged image back through Stable Diffusion at low denoise (0.2 to 0.4) so the model paints in new texture. The usual pipeline uses both: an ESRGAN model to enlarge, then a diffusion pass to add detail. Ultimate SD Upscale does exactly this, using your chosen ESRGAN model as its upscaler and refining each tile with diffusion. Full workflow in how to upscale NSFW AI images.
# Typical Ultimate SD Upscale pairing
Upscaler model: 4x-UltraSharp # enlarges
Denoise: 0.3 # diffusion adds detail
Tile size: 512
Steps: 20Worked example: choosing a model
Say you have a photoreal 896×1152 render of an adult character and you want a 4K wallpaper.
- The face is the priority, so reach for Remacri for natural skin or UltraSharp for extra crispness.
- Load it into Ultimate SD Upscale as the upscaler model, target 4x.
- Set denoise 0.3 so the diffusion pass adds pores and hair detail without changing the face.
- Tile at 512 to fit VRAM. See low VRAM upscaling.
- Export PNG, review at 100 percent.
If the same render were anime, you would swap the upscaler model to 4x-AnimeSharp and lower denoise slightly to keep flat color clean. Everything else stays the same.
Try our free NSFW generator to make base images, then upscale them with the model that matches your style.
Common mistakes
- Wrong folder. Models in the wrong directory never appear. Double-check models/ESRGAN versus models/upscale_models for your app.
- Using one model for everything. Photoreal and anime need different model families. Keep several installed.
- Expecting added detail from ESRGAN alone. It enlarges and sharpens only. Add a diffusion pass for real detail.
- Over-sharpening with an aggressive model on a sharp source. Switch to Remacri or add diffusion at low denoise.
- Downloading the wrong file. Grab the .pth or .safetensors weights, not training data or a sample image.
Photoreal vs anime, quick reference
- Photoreal portraits: Remacri (natural) or UltraSharp (crisp).
- Photoreal general: NMKD-Siax or RealESRGAN_x4plus.
- Noisy or compressed photos: SwinIR.
- Anime flat color: RealESRGAN_x4plus_anime_6B.
- Anime crisp line art: 4x-AnimeSharp.
A note on safety
These models process whatever you feed them, so the responsibility is in your input. Upscale only adult, fictional, AI-generated or fully owned subjects, and keep a baseline safety negative when you generate the source:
negative: child, minor, underage, loli, shota, lowres, deformedA clean source plus the right model plus a local pipeline gives censorship-free, private, high-quality results. Generate clean base images for free and keep everything on your own machine.
Understanding the architectures
Not all upscaler models share the same underlying architecture, and it affects how they behave.
- ESRGAN and Real-ESRGAN are the classic GAN-based architectures. Fast, sharp, and the basis for most community models (UltraSharp, Remacri, Siax, AnimeSharp). They occasionally produce GAN artifacts like fine grid patterns on flat areas.
- SwinIR and HAT are transformer-based. Slower but smoother, with better noise handling and fewer hard artifacts. Good for restoration-leaning jobs.
- DAT and SPAN are newer architectures appearing on OpenModelDB that aim to beat ESRGAN on detail retention. Worth testing if you want the latest.
For day-to-day adult work, the ESRGAN-family community models cover almost everything. Transformer models are a nice tool to keep around for noisy sources.
How to test a new model quickly
When you grab a model from OpenModelDB, do not trust the sample images blindly, since they are cherry-picked. Run your own test:
# Quick A/B test in chaiNNer or Upscayl
1. Pick one representative crop of YOUR art (a face, some skin, some fabric)
2. Upscale 4x with candidate model A
3. Upscale 4x with candidate model B
4. Compare at 100 percent: halos? plastic skin? clean lines?
5. Keep the winner, delete the restA single face-and-skin crop tells you in 30 seconds whether a model suits your style. This beats reading a dozen reviews.
Pairing models with diffusion: a worked recipe
The strongest pipeline uses an ESRGAN model as the enlarger inside a diffusion upscale. The model does the clean 4x, the diffusion pass adds the detail the model cannot.
# Recommended pairing for photoreal adult art
Stage 1 enlarge: 4x-UltraSharp (or Remacri for softer skin)
Stage 2 refine: img2img / Ultimate SD Upscale
denoise: 0.3
steps: 20
prompt: reuse the original generation prompt
negative: child, minor, underage, loli, shota, lowres, deformedFor anime swap stage 1 to 4x-AnimeSharp and drop denoise to 0.25. This recipe is the heart of how to upscale NSFW AI images and works in both ComfyUI and Automatic1111.
When deterministic alone is enough
You do not always need a diffusion pass. If your source is already sharp and detailed and you only need more pixels (for example a clean 1536px render going to a 3072px print), a straight ESRGAN upscale with UltraSharp or Siax is faster, cheaper in VRAM, and perfectly good. Add the diffusion pass only when the source lacks the fine detail you want, such as smooth skin that should have pores. Knowing when to skip the diffusion step saves a lot of time. Try our free NSFW generator to build sharp sources that upscale cleanly with deterministic models alone.
Specialized and niche models worth knowing
Beyond the staples, a few models solve specific problems:
- 4x_NMKD-Superscale is detail-heavy and excels on fabric, leather, and surface texture, useful when clothing detail matters.
- 4x_Nickelback_70000G and similar community photoreal models trade a little sharpness for very natural skin, an alternative to Remacri.
- 1x models exist too, which do not change resolution but clean up JPEG artifacts or denoise. Run a 1x cleanup before a 4x upscale on a noisy source.
- 4x_LSDIR and DAT-based models are newer high-detail options on OpenModelDB worth testing against UltraSharp.
You do not need all of these. Start with the core five, then add a specialist only when a specific job calls for it. Hoarding models just clutters your dropdown.
Model file formats and safety
Upscaler models come as .pth (older, pickle-based) or .safetensors (newer, safe format). Prefer .safetensors when available because .pth files can in theory carry malicious code, though models from OpenModelDB and official Hugging Face repos are trustworthy. Either format loads the same way in ComfyUI, Automatic1111, chaiNNer, and Upscayl. File sizes are small, so keeping a curated folder of five to eight models costs almost no disk space.
How models interact with your checkpoint
A point that confuses newcomers: the upscaler model and your generation checkpoint are separate things. The checkpoint (for example a photoreal checkpoint or Illustrious) generates and, in a diffusion upscale, refines the image. The upscaler model only enlarges. In Ultimate SD Upscale you set both: the checkpoint does the refine pass, the ESRGAN model does the enlargement. Keeping these roles straight is the key to understanding why your results look the way they do. If the enlargement is harsh, change the ESRGAN model; if the refined detail looks off-style, change the checkpoint or denoise.
Keeping your model library organized
With a handful of models installed it helps to name and group them clearly. Many people prefix files so the dropdown sorts sensibly, for example putting photoreal models together and anime models together. Keep a short text note of which model you prefer for which job (portraits, full scenes, anime, noisy sources) so you are not re-testing every session. A tidy, curated library of five to eight trusted models beats a sprawling folder of forty you never use, and it makes the dropdown fast to navigate mid-workflow. When a genuinely better model appears on OpenModelDB, test it against your current favorite on your own crops and only then replace the incumbent.
Bottom line
Keep three to four models installed: UltraSharp and Remacri for photoreal, the anime 6B and AnimeSharp for 2D, and NMKD-Siax as a flexible default. Get them from OpenModelDB or Hugging Face, drop them in the right folder, match the family to your art, and remember that for genuine added detail you pair the ESRGAN model with a low-denoise diffusion pass. Test new models on your own crops, not their sample images, and only add a diffusion pass when the source genuinely lacks detail.
Frequently asked questions
What is the best upscaler model for photoreal NSFW art?
4x-UltraSharp for crisp results and 4x_foolhardy_Remacri for natural skin are the top two. Remacri is softer with less haloing, which makes it the favorite for portraits, while UltraSharp is sharper for general photoreal work. Many users keep both installed and switch based on the image.
Where do I put upscaler models in ComfyUI and Automatic1111?
In ComfyUI place them in ComfyUI/models/upscale_models. In Automatic1111 and Forge use stable-diffusion-webui/models/ESRGAN. Drop in the .pth or .safetensors file, then refresh or restart so the model appears in the dropdown or node. Wrong folder is the most common reason a model does not show up.
Where can I download ESRGAN upscaler models?
OpenModelDB (openmodeldb.info) is the community catalog with samples and use cases, and Hugging Face hosts many of the same models including official Real-ESRGAN releases. Download the model weights file, typically 60MB to 70MB for a 4x model, not training data or sample images.
Do upscaler models add new detail to images?
No. ESRGAN-family models are deterministic. They enlarge and sharpen existing detail but invent nothing new, so a soft source stays soft, just bigger. To genuinely add detail you pair them with a low-denoise diffusion pass, as Ultimate SD Upscale does.
Which upscaler model is best for anime NSFW art?
RealESRGAN_x4plus_anime_6B is the official lightweight choice that keeps flat color flat and lines clean. 4x-AnimeSharp is sharper and more vivid for crisp line art. Never use a photoreal model on anime, since it adds unwanted texture to flat color areas.
What is the difference between UltraSharp and Remacri?
UltraSharp is sharper and crisper with stronger edge definition, but it can produce slight halos on already-sharp sources. Remacri is softer and more natural with minimal haloing, which makes it better for skin and portraits. Choose based on whether your source needs more crispness or more softness.
Can I use these upscaler models on a low VRAM GPU?
Yes. Deterministic ESRGAN upscaling is light and runs fine on 4GB to 6GB. The VRAM cost comes from pairing them with a diffusion pass at high resolution, which you solve with tiling in Ultimate SD Upscale using small 512-pixel tiles.
Why does my upscaler model not appear in the dropdown?
Almost always because the file is in the wrong folder or the app was not refreshed. Confirm the path matches your app (models/upscale_models for ComfyUI, models/ESRGAN for Automatic1111 and Forge), make sure you downloaded the weights file, then restart the app.
