You have a good image. But it's too small to print, too soft to post, or too compressed to use professionally.
You don't want to just stretch pixels. You want the detail to actually be there.
Low-res image in. Sharp, detailed 2K output out.
Run it now on Floyo!
Why This Workflow Is Different
Most upscalers enlarge the image and call it done. Pixels get bigger. Detail doesn't improve.
This workflow runs two stages. RealESRGAN x2 handles the raw upscale recovering edges and textures with a dedicated super-resolution model. Then Z-Image Turbo runs a guided diffusion pass at low denoise (0.33) to refine and sharpen detail without changing the composition. You get an image that's not just bigger, but genuinely more detailed.
two-stage pipeline: super-resolution then diffusion refinement
only 5 sampling steps fast without sacrificing quality
prompt-guided refinement via Qwen 3 4B text encoder
composition stays intact throughout
generates in about 42 seconds
How It Works
Stage 1: RealESRGAN x2 performs the initial upscale. It's a dedicated super-resolution model trained to recover fine texture, edges, and surface detail from low-resolution inputs. Not just enlarging reconstructing.
Stage 2: Z-Image Turbo runs a KSampler pass at denoise 0.33 on the upscaled image. Low enough to preserve the composition entirely. Just enough to add realistic texture and clarity that pure upscaling misses. 5 steps with DPM++ 2M SDE. Done in under a minute.
Qwen 3 4B: acts as the text encoder. You can guide the refinement direction with a prompt useful when you want to steer the enhancement toward a specific look or detail style.
Key Inputs
Your Image
Any image you want upscaled. The workflow handles the resolution scaling automatically.
Works well with:
AI-generated images that need a final quality pass
low-resolution photos and scans
product images for e-commerce or print
game screenshots and concept art
compressed or degraded images
Works less well with:
extremely low-resolution inputs under 128px too little information to recover from
heavily motion-blurred images where no detail exists to reconstruct
images where you need strict pixel-perfect preservation (the diffusion pass adds refinement, not pure upscaling)
Prompt (optional)
Guide the refinement direction. Leave it blank for neutral enhancement, or add style descriptors to steer the output.
Examples:
"sharp detail, photorealistic, clean edges"neutral enhancement push"natural skin texture, sharp eyes, professional photography"portrait refinement"sharp product details, clean background, e-commerce photography"product focus"detailed fur texture, sharp edges, wildlife photography"nature/animal shots
Denoise (default: 0.33)
This is set low intentionally. At 0.33 the diffusion pass refines without altering structure. Raising it changes the image more significantly, useful if you also want to shift style, but will move further from the original.
What This Is Great For
AI image finishing: Run your generations through this as a final pass before using or sharing. Sharpens micro-detail and removes the slightly digital look common in AI outputs.
Photo restoration and recovery: Recover detail from old, compressed, or low-resolution photos before printing or archiving.
E-commerce and product photography: Upscale product images to print or display resolution without losing sharpness on edges and textures.
Print preparation: Take any image to 2K resolution for large-format printing, posters, or editorial use.
Concept art and illustration: Sharpen renders and illustrations for portfolio use or client delivery.
What to Watch Out For
Denoise above 0.5 starts to alter the image noticeably. Use 0.33 for pure enhancement. Only raise it if you also want to shift style or correct something in the original.
Very degraded inputs have a ceiling. The workflow reconstructs detail based on what exists in the image. A heavily compressed or blurry source produces a sharper version of a limited original — not a fully recovered image. Start with the best source you have.
Prompt guidance is subtle at low denoise. Don't expect dramatic style changes from the prompt at 0.33. It nudges the refinement direction, not the content.