Skip to content
DreamSprite Logo DreamSprite

Upscaling & Refinement

Most ComfyUI workflows start with lower resolutions (512x768, 768x1024, or 1024x1024) because generating directly at 4K+ eats VRAM and takes forever. Upscaling turns those base images into crisp, high-res final assets without losing quality - and refinement fixes flaws while adding detail.

There are two main phases:

  • Upscaling: Increase resolution (e.g., 2x, 4x) while preserving or enhancing sharpness.
  • Refinement: Iteratively improve details, fix artifacts, add texture via low-denoise img2img passes.

Why Upscale & Refine?

  • Base gens often look good but soft/blurry at higher zoom.
  • High-res reveals flaws (hands, faces, edges).
  • Refinement adds realism, skin texture, fine details that low-step gens miss.

Here are the most used and effective techniques:

1. Simple Latent Upscale (Fastest, Built-in)

  • Use Latent Upscale or Upscale Image (Latent) nodes.
  • Scale by 1.5-2x in latent space, then VAE Decode.
  • Pros: Very fast, low VRAM.
  • Cons: Can be soft/blurry; best as a quick first step.
  • Tip: Follow with a small img2img pass (denoise 0.3-0.5) to sharpen.

2. Model-Based Upscalers (ESRGAN, Real-ESRGAN, SwinIR, 4x-UltraSharp)

  • Load models via Upscale Image (using Model) node.
  • Popular models (download from Civitai or OpenModelDB):
    • 4x-UltraSharp: Excellent general-purpose 4x upscale - sharp, preserves details.
    • 4x_NMKD-Superscale: Great for realistic/photographic.
    • Real-ESRGAN x4plus / Anime variants: Strong for photos or anime.
    • SwinIR / Swin2SR: Transformer-based, excellent restoration and sharpness.
  • Pros: Fast, high quality, no diffusion needed.
  • Cons: Can over-sharpen or add halos if mismatched to content.
  • Tip: Use for 2x-4x jumps after latent upscale.

3. Diffusion-Based / Tiled Upscaling (Ultimate SD Upscale)

  • The gold standard for high-quality, creative upscales in 2026.
  • Install ComfyUI_UltimateSDUpscale custom nodes via Manager.
  • Workflow: Tile the image, run low-denoise img2img on each tile, stitch back.
  • Key settings:
    • Upscale by: 2-4x
    • Denoise: 0.25-0.5 (low for fidelity, higher for more detail invention)
    • Tile size: 512-1024 (balance quality vs VRAM)
    • ControlNet Tile or Tile Resample for edge consistency
  • Pros: Adds real detail/texture (e.g., skin pores, fabric weave), fixes artifacts.
  • Cons: Slower, more VRAM.
  • Variants: SeedVR2, SUPIR, or Flux-based upscalers for even better results (if you have the models).

4. Iterative Img2Img Refinement

  • After upscaling, run multiple low-denoise passes to polish.
  • Typical chain:
    1. Base image -> upscale 2x (latent or model).
    2. Img2img pass 1: Denoise 0.4-0.6, add “highly detailed, sharp focus” prompt.
    3. Img2img pass 2: Denoise 0.2-0.3, focus on “refine details, better anatomy”.
    4. Final pass: Denoise 0.1-0.2 for subtle enhancements.
  • Pros: Gradually builds quality, fixes specific flaws (hands, faces).
  • Cons: Takes longer - best for final touches.

5. Face-Specific Refinement (FaceDetailer, Reactor)

  • Use FaceDetailer or Reactor nodes (from custom packs like ComfyUI Impact Pack).
  • Detects faces, runs targeted img2img or restoration (GFPGAN/CodeFormer).
  • Great for portraits where faces look off after upscaling.

Quick Tips & Best Practices

  • Order: Generate base (low res) -> ControlNet/pose lock -> Sampler -> Upscale -> Refinement passes.
  • Denoise sweet spot: 0.3-0.5 for most upscales/refinements - higher adds creativity, lower preserves original.
  • ControlNet in upscaling: Add Tile or Lineart ControlNet (strength 0.4-0.7, late timing) to keep edges sharp.
  • VRAM saving: Use tiled methods (Ultimate SD Upscale) or lower tile size.
  • Test small: Always upscale a crop first.
  • Save workflows: Once you have a good chain (e.g., Ultimate SD + FaceDetailer), save as JSON for reuse.

Upscaling & refinement turns “good enough” into “professional” - it’s where the wow factor happens.