AI Image Upscaling Locally: Enhance Stable Diffusion Output in 2026
How to upscale AI-generated images locally using Real-ESRGAN, Ultimate SD Upscale, and Stable Diffusion tiled upscalers. Get 4K+ resolution from your local uncensored generations — completely offline.
Why Upscale?
SDXL generates at 1024×1024. Flux outputs at similar resolutions. That's fine for web use, but for printing, wallpapers, or professional work, you need 4K or higher.
AI upscaling doesn't just resize — it regenerates detail. Skin pores, fabric textures, background elements that were blurry at 1024px become sharp and defined at 4096px. All locally, all offline.
Method 1: Real-ESRGAN (Quick & Easy)
Built into Forge UI. Best for fast, simple upscaling without regenerating details.
- Go to the Extras tab in Forge UI
- Upload your image or send from txt2img
- Select "R-ESRGAN 4x+" as the upscaler
- Set resize to 2× or 4×
- Hit Generate — takes 2–10 seconds
Best upscaler models:
- R-ESRGAN 4x+ — best all-around for photorealistic
- R-ESRGAN 4x+ Anime6B — optimized for anime/illustration
- 4x-UltraSharp — community favorite for maximum sharpness
Method 2: Ultimate SD Upscale (Best Quality)
This extension tiles the image and regenerates each tile with Stable Diffusion, adding real detail that didn't exist at the original resolution.
- Install the extension in Forge: Extensions → Install from URL →
https://github.com/Coyote-A/ultimate-upscale-for-automatic1111 - In img2img, upload your image
- Set denoising to 0.2–0.4 (lower = closer to original, higher = more creative)
- Select "Ultimate SD Upscale" in the Script dropdown
- Set upscaler to R-ESRGAN 4x+, tile size 512 or 768
- Generate — takes 1–5 minutes depending on target size
Pro tip: Use the same prompt and seed from the original generation for most consistent results.
Method 3: ControlNet Tile (Most Control)
For maximum quality, combine ControlNet Tile with img2img:
- Pre-upscale with Real-ESRGAN to your target resolution first
- Send to img2img at the upscaled resolution
- Enable ControlNet Tile with the pre-upscaled image as input
- Set denoising 0.3–0.5 — ControlNet Tile preserves overall composition while SD adds detail
This produces the highest quality upscales but requires more VRAM (12+ GB recommended) and takes longer.
Installing upscale extensions manually? That means cloning Git repos, matching Python versions, downloading upscaler models from three different sources, and hoping nothing conflicts. Most people hit errors on the first try. LocalForge AI includes Real-ESRGAN and Ultimate SD Upscale pre-installed — generate and upscale in the same session, zero additional setup.
VRAM Requirements for Upscaling
- Real-ESRGAN only: 2 GB VRAM — works on any modern GPU
- Ultimate SD Upscale (SDXL): 8+ GB for 2× upscale, 12+ GB for 4×
- ControlNet Tile + img2img: 10+ GB for 2×, 16+ GB for 4× at high resolution
If VRAM is tight, use smaller tile sizes (256 or 512) in Ultimate SD Upscale — it processes tiles sequentially, using less peak VRAM at the cost of more processing time.
FAQ
Can I upscale images that weren't made with SD?
Yes. Real-ESRGAN works on any image — photos, screenshots, artwork. The SD-based methods (Ultimate Upscale, ControlNet Tile) work best on AI-generated images where the model can coherently add detail.
What's the maximum resolution I can achieve?
Theoretically unlimited with tiled upscaling. Practically, 4096×4096 to 8192×8192 is the sweet spot. Beyond that, generation times become very long and the visual benefit diminishes.