Concept · Open source vs API

Real-ESRGAN API vs Self-Hosted

Real-ESRGAN is the open-source upscaling standard: excellent quality, free weights, and a non-trivial service layer around it. This is the honest breakdown of what self-hosting really costs vs calling an API.

Real-ESRGANSelf-hostingUpscaling

What Real-ESRGAN is

Real-ESRGAN is the most popular open-source GAN upscaler, with models for faces (GFPGAN-integrated) and general imagery. Quality is genuinely excellent, which is why it is the default baseline for self-hosted upscaling — and why our API is benchmarked against it.

The self-hosting cost breakdown

Self-hosted Real-ESRGANmlslabs API
GPU1× 8–12 GB VRAM per lane; idle cost between jobs$0 — workers scale to your queue
SpeedFast on GPU; ~10–60 s/min of 720p videoParallel workers, similar per-frame speed
OpsCUDA env, model files, queue, retries, monitoringNone
Temporal coherenceYou add it (Real-ESRGAN is per-frame)Built-in temporal smoothing
Model updatesYou track themShipped continuously
Cost shapeFixed infra + your timePer minute of video

The per-frame problem nobody mentions

Stock Real-ESRGAN upscales frame by frame — each frame independent. On video, that causes flicker: texture that changes between frames. Production pipelines add temporal consistency (optical-flow guidance, blending) — the part of the work that is invisible in demos and essential in deliverables.

When self-hosting wins

  • Data must never leave your environment
  • GPU fleet already exists with spare capacity
  • Full control over every model and setting
  • Deep customization of the pipeline

When the API wins

  • Batch libraries with variable volume
  • No GPU ops team
  • Temporal coherence out of the box
  • Webhook-driven pipelines
  • Pay-per-use vs fixed infra

FAQ

Common questions

Is the API just Real-ESRGAN in the cloud?

The model family is comparable; the API adds temporal coherence, per-content tuning and the whole service layer (queue, retries, webhooks) that self-hosting leaves to you.

How much GPU do I need to self-host?

8–12 GB VRAM for comfortable 1080p work. Upscaling to 4K at batch scale wants a bigger card or more lanes.

Can I migrate from self-hosted to API?

Yes — same inputs/outputs (object storage), same idea, fewer machines. The tutorial pages walk through the swap.