NSFW AI Hires Fix 2026: Complete Guide

NSFW hires fix in AUTOMATIC1111 in 2026: enable hires fix in txt2img, set upscaler to 4x-UltraSharp (realistic) or RealESRGAN x4-Anime6B (anime), upscale 1.5x to 2x, denoising 0.35-0.5, hires steps 20-25. This avoids the doubled-anatomy bug from direct high-resolution generation while producing sharp NSFW output.

Hires fix is the two-stage generation pipeline built into AUTOMATIC1111 (and replicated as a node pattern in ComfyUI). It exists because diffusion models trained at one resolution produce broken output when asked to generate directly at a different resolution. The classic symptom: doubled anatomy. Generate at 1024×1536 on a model trained for 512×768 and you get a result with two heads, three arms, or two stacked torsos, because the model treats the large canvas as multiple instances of what it knows how to produce.

Hires fix solves this by separating the resolution-bound parts of generation from the resolution-flexible parts. The base pass at native resolution produces a coherent composition. The upscaler model resizes the image. The partial diffusion pass at the higher resolution refines detail without rebuilding the composition. The result is high-resolution NSFW output with intact anatomy.

This guide covers the complete hires fix workflow for NSFW work in 2026: settings that work, the doubled-anatomy bug and how to avoid it, hires fix in SDXL versus SD1.5, and the ComfyUI equivalent for users who have moved away from A1111.

The Hires Fix Settings That Actually Matter

Upscaler. The model used for the resize step. 4x-UltraSharp for realistic content, RealESRGAN x4-Anime6B for anime. SwinIR as backup. Avoid Latent upscaler for NSFW; it produces good results but is unpredictable on anatomical detail.

Upscale by. The scaling factor. 1.5x is the safe default that rarely introduces artifacts. 2x is fine for SDXL going from 1024 to 2048. 3x+ in a single hires fix pass tends to produce artifacts; for bigger scales use iterative img2img passes.

Hires steps. The number of diffusion steps in the refinement pass. 0 means “use the same step count as the base pass.” 20-25 explicit steps usually produces the cleanest output. More steps rarely helps and is just slower.

Denoising strength. The most critical setting. Controls how much of the upscaled image gets regenerated. 0.35-0.5 is the working range for NSFW. Below 0.3 the diffusion pass has nothing to do; above 0.55 the composition starts drifting and you risk anatomy changes.

Hires sampler. Defaults to “Use same sampler.” This is usually correct. Mixing samplers between base and hires can produce inconsistent rendering.

The Doubled-Anatomy Bug

If you generate at resolutions far from the model’s training resolution without hires fix, you get composition errors. The most common is doubled anatomy: two heads, three arms, two torsos. This is because diffusion models, especially older SD1.5 derivatives, treat each “training-resolution-sized region” of a too-large canvas as an independent generation. The model places a person in each region and then tries to merge them, often unsuccessfully.

Three preventions. First, generate at the model’s training resolution (512×768 for SD1.5, 1024×1024 for SDXL) and use hires fix to scale up. Second, even with hires fix, don’t ask for a base resolution far from training resolution. Generating at 768×1280 on SD1.5 is borderline; 1024×1536 is asking for trouble. Third, if doubled anatomy appears in your hires fix output (not just the base), lower hires denoising to 0.3-0.35 so the refinement pass doesn’t re-introduce the issue.

Hires Fix for SD1.5 NSFW

SD1.5 trained at 512×512. Practical base resolutions: 512×512 square, 512×768 portrait, 768×512 landscape. Hires fix from these to 1.5x or 2x produces 768×1152 to 1024×1536 final resolutions. For NSFW SD1.5 work (Counterfeit, MeinaMix, AbyssOrangeMix), the standard settings are: base 512×768, hires fix 4x-UltraSharp at 1.5x, hires steps 20, denoising 0.45. This produces 768×1152 final with sharp detail and intact anatomy.

Hires Fix for SDXL NSFW

SDXL trained at 1024×1024 with aspect ratio buckets (832×1216, 1216×832, etc.). Practical base resolutions stay in those buckets. Hires fix from 1024×1024 to 1.5x produces 1536×1536; to 2x produces 2048×2048. SDXL hires fix settings: 4x-UltraSharp at 1.5x or 2x, hires steps 20-25, denoising 0.4. SDXL’s larger native resolution means you often need less aggressive upscaling than SD1.5 to reach a useful final size.

SDXL also supports the “Refiner” model which is a different but related concept: a second SDXL-base-companion model that refines the last 20% of generation steps. Refiner is not hires fix; both can be used together. For most NSFW SDXL work, hires fix alone is sufficient and the refiner adds little value at the cost of complexity.

Hires Fix Equivalent in ComfyUI

ComfyUI doesn’t have a single hires fix checkbox; you build the equivalent workflow with nodes. First KSampler at base resolution. VAE Decode to pixels. Upscale Image (using Model) with 4x-UltraSharp. VAE Encode back to latent at the new resolution. Second KSampler with low denoising (0.4) using the upscaled latent as input. VAE Decode final. Save Image.

The ComfyUI version gives more control: you can swap upscalers between the pixel-domain and latent-domain stages, use different prompts on the hires pass (sometimes useful to add detail-focused tags like “ultra-detailed” only at hires), or chain multiple hires passes. For most users, the A1111 hires fix checkbox is faster; for advanced control, the ComfyUI workflow earns its complexity.

Settings Cheat Sheet

Common Hires Fix Problems

Doubled anatomy after hires fix. Cause: hires denoising too high (0.55+) or base resolution too far from training. Drop denoising to 0.35 and ensure base resolution matches training. Face still blurry after hires fix. Hires fix is broad refinement, not face-specific. Use ADetailer extension after hires fix for face refinement. Composition changed unexpectedly. Denoising is too high; the refinement pass is doing more than refinement. Lower to 0.4. Artifacts in skin texture. Wrong upscaler for content type, or upscale factor too high. Switch upscaler or reduce factor. VRAM error during hires fix. Reduce upscale factor or generate at a smaller base resolution. SDXL hires fix at 2x can exceed 12GB VRAM.

Recommended Defaults

SD1.5 NSFW: base 512×768, hires fix 4x-UltraSharp at 1.5x, denoising 0.45, hires steps 20. SDXL NSFW: base 1024×1024 (or appropriate aspect bucket), hires fix 4x-UltraSharp at 1.5x, denoising 0.4, hires steps 25. Anime variants: swap the upscaler to RealESRGAN x4-Anime6B and drop denoising by 0.05. These settings cover the majority of NSFW use cases without tuning.

For pure upscaling workflows (no diffusion refinement) see our NSFW AI upscaler complete guide. For face-specific refinement after hires fix, see our ADetailer guide. For ComfyUI-based equivalents, our ComfyUI guide covers the full workflow. For local Stable Diffusion setup, see best local NSFW AI image generator.

Picking the Right Upscaler for Hires Fix

The upscaler dropdown in the hires fix panel matters more than most users realize, and the right choice depends on the content type. For realistic NSFW, 4x-UltraSharp is the dependable default: it adds crisp skin and fabric detail without introducing the artifacts that can appear with more aggressive models. R-ESRGAN 4x+ is a slightly softer, neutral alternative that suits portraits where you do not want the texture pushed hard. For anime and 2D NSFW, R-ESRGAN 4x+ Anime6B or 4x-AnimeSharp are the correct picks because they keep flat color clean and line art crisp, whereas a photoreal upscaler tends to add unwanted noise and texture to cel-shaded art. The Latent upscaler behaves differently from all of these: it upscales in latent space rather than pixel space, which gives the diffusion pass more freedom to add detail but requires a higher denoising strength (0.5 to 0.7) to avoid a blurry result, and it can change the image more. Use a pixel-space upscaler (the UltraSharp or ESRGAN family) when you want the composition preserved, and reserve Latent for when you actively want hires fix to refine and add detail. Match the upscaler to the checkpoint style and the result improves immediately.

A Deep Fix for the Doubled-Anatomy Bug

The doubled-anatomy bug, where hires fix produces a second head, extra limbs, or duplicated body parts, is the most common hires complaint, and it has a clear root cause: the denoising strength on the hires pass is too high for the upscale factor, so the model effectively treats the enlarged canvas as new space to fill and redraws content that should have stayed singular. The first and most effective fix is lowering hires denoising to the 0.3 to 0.45 band; at this strength the pass refines what exists instead of inventing new anatomy. The second factor is the upscale multiplier: large jumps like 2.5x or 3x in a single hires pass strongly encourage duplication, so cap the hires upscale at around 1.5x to 2x and reach higher resolutions through a separate img2img or tiled upscale pass instead. Third, reinforce the negative prompt with multiple bodies, extra heads, duplicate, extra limbs, cloned figure, since the hires pass reads the negative prompt and these terms suppress duplication directly. Fourth, very high hires step counts give the model more opportunity to drift, so keep hires steps modest. When a doubled result still slips through, the fastest recovery is to inpaint the duplicated region away at moderate denoising rather than rerolling the whole image; our inpainting workflow guide covers the masking approach for removing a duplicated limb or head cleanly.

The Hires Steps and Denoising Matrix

Hires steps and denoising strength interact, and thinking of them as a small matrix removes most of the guesswork. With low denoising (0.25 to 0.35), low to moderate hires steps (10 to 20) are enough, because the pass is only cleaning and sharpening; this combination is the safest and the right starting point for a finished image you simply want larger and crisper. With moderate denoising (0.4 to 0.5), use moderate hires steps (20 to 30) so the pass has room to add genuine detail like skin texture and hair strands; this is the band for a soft source image that needs enrichment, accepting minor drift. High denoising (0.55 plus) combined with high hires steps effectively regenerates the image and is rarely what you want from hires fix, since at that point a dedicated img2img or upscale workflow gives better control. The two combinations to avoid are high denoising with low steps, which produces a noisy half-finished look, and high denoising with a large upscale factor, which triggers the doubled-anatomy bug. For most NSFW work, set hires denoising around 0.35, hires steps around 15 to 20, and upscale 1.5x to 2x, then adjust only if the result is too soft (raise denoising) or showing duplication (lower denoising and upscale factor).

Final Notes on Hires Fix

Hires fix is the single highest-impact setting most NSFW generators leave misconfigured. The default denoising value is almost always too high, which is what produces the doubled-anatomy bug where a second torso, extra limbs, or duplicated faces appear in the upscaled output. Drop denoising to the 0.3 to 0.45 range and the bug largely disappears. The second common mistake is pairing hires fix with too few hires steps, which leaves the upscaled image soft. Match your upscaler to your content type, keep denoising conservative, and use enough hires steps to actually resolve detail. For anatomy errors that survive a properly configured hires pass, a targeted inpainting cleanup using our inpainting workflow is faster than re-rolling the whole generation.

Hires Fix is a built-in feature of Stable Diffusion WebUI. For the upscaler models it depends on, see the model repositories hosted on Hugging Face.

Frequently Asked Questions

What is hires fix in AUTOMATIC1111?

Hires fix is a two-stage generation pipeline in AUTOMATIC1111: generate at base resolution, then upscale with a chosen upscaler model, then run a partial diffusion pass at the higher resolution. The result has more detail than a pure upscale because the diffusion pass synthesizes fine features. It is the standard way to get high-resolution NSFW outputs without the doubled-anatomy bug from direct high-resolution generation.

Why not just generate at high resolution directly?

Stable Diffusion models are trained at fixed resolutions (512×512 for SD1.5, 1024×1024 for SDXL). Generating directly at higher resolutions produces the doubled-anatomy bug: two heads, three arms, extra torsos, because the model treats the larger canvas as multiple instances of its training data. Hires fix solves this by generating at native resolution first and then refining at higher resolution.

What hires fix settings work for NSFW?

Upscaler: 4x-UltraSharp for realistic, RealESRGAN x4-Anime6B for anime. Upscale factor: 1.5x or 2x. Denoising: 0.35-0.5 (lower preserves more of the base image; higher adds more detail but risks composition drift). Hires steps: 20-25 typically sufficient. Match the hires sampler to your txt2img sampler.

What denoising strength should I use for hires fix?

0.35-0.5 is the working range. 0.35 keeps composition almost identical, adds sharpness only. 0.45 is a balanced default. 0.5+ introduces creative variation and may shift anatomy. For NSFW where preserving the base composition matters (you got a good pose, don’t break it), stay in 0.35-0.45.

Does hires fix work with SDXL?

Yes. SDXL is trained at 1024×1024 so hires fix from there to 1536×1536 or 2048×2048 is the standard SDXL workflow. Settings are similar to SD1.5. Some SDXL workflows skip hires fix in favor of img2img upscaling at smaller increments (1.25x at a time) to avoid VRAM issues.

How do I avoid the doubled anatomy bug?

Three rules. Generate at the model’s native resolution (512×768 for SD1.5, 1024×1024 for SDXL). Use hires fix to scale up rather than going direct. Keep your txt2img resolution close to the model’s training aspect ratios (don’t generate at 1024×2048 on SD1.5). If doubled anatomy still appears in hires fix output, lower denoising to 0.3-0.35.

Does ComfyUI have an equivalent of hires fix?

Yes, but it is not a single setting. In ComfyUI you build the equivalent workflow: first KSampler at base resolution, Upscale Image (using Model), second KSampler at higher resolution with low denoising. This is more setup than A1111’s checkbox but gives precise control over each stage.

Should I use hires fix or upscale-then-img2img?

Hires fix is the cleaner one-pass option for going from base resolution to roughly 2x. For larger upscales (3x, 4x), iterative img2img at smaller increments (1.5x at a time) produces fewer artifacts than one big hires fix pass. For most NSFW work, hires fix at 1.5x or 2x is sufficient.