Real-ESRGAN
The default OSS upscaler, still.
Open Source 2–6 GB VRAM
- Min VRAM
- 2 GB
- GPU class
- Entry GPU
- Quant
- FP16
Actually FreeNo SignupOpen SourceWatermark-Free
SUPIR
Diffusion-based photorealistic upscaler.
Open Source 12–24 GB VRAMRuns locally
Actually FreeNo SignupWatermark-FreePlugin
Requires: ComfyUI
Visit SUPIRUpdated 2025-12-08 · Direct link
Hardware requirements
Runs locally · Mid GPU (12 GB)
Min VRAM
12 GB
Rec. VRAM
24 GB
Min RAM
16 GB
Rec. RAM
32 GB
Disk
20 GB
GPU class
Mid GPU
11.8+No Apple SiliconGPU RequiredQuant: FP16
FP16 fits in 12 GB with tiling; 24 GB lets you process larger inputs.
Screenshot placeholder · SUPIR
What is SUPIR?
SUPIR (Scaling Up to Photorealistic Image Restoration) uses an SDXL backbone for upscaling — restoring fine detail by literally redrawing it instead of interpolating. Heavier than Real-ESRGAN by an order of magnitude but the photorealistic detail recovery on real photos is in a different class.
Pros & cons
Pros
- Reconstructs detail rather than interpolating — best for photos
- Strong prompt conditioning for guided restoration
- ComfyUI nodes available — fits straight into existing pipelines
Cons
- 16 GB VRAM minimum, 24 GB+ comfortable
- Slow vs. Real-ESRGAN — minutes vs. seconds
- Non-commercial licence
What's actually free?
Non-commercial research licence; weights & code public.
✓ Actually FreeNo SignupWatermark-Free
Alternatives
Topaz Video AI
GPU-accelerated upscaling, frame-interp, denoise.
Paid · from $299 4–12 GB VRAM
- Min VRAM
- 4 GB
- GPU class
- Mid GPU
- Quant
- —
Watermark-FreeHobbyist-Friendly