Nodes/ComfyUI-RealRestorer/RealRestorer Sampler
ComfyUI Node

RealRestorer Sampler

Restores a degraded image using RealRestorer. Select a Task Preset for common restoration tasks, or choose 'Custom' and write your own instruction. Paper defaults: 28 steps, guidance 3.0, size 1024, seed 42.

By yfyang007·Created 5 months ago·Updated 4 months ago· 12
RealRestorer Sampler
  • RR_Model
  • image
  • image
task_presetGeneral Restore
instruction
seed42
steps28
guidance_scale3.0
size_level1024
device_strategyauto
CategoryRealRestorer

Inputs (9)

NameTypeDefaultDescription
RR_ModelRR_MODEL
imageIMAGE
task_presetCOMBOGeneral RestoreSelect a restoration task. Each preset uses the prompt recommended in the RealRestorer paper for that degradation type. Choose 'Custom' to write your own instruction below.
instructionSTRINGCustom restoration instruction. Only used when Task Preset is set to 'Custom'. For all other presets, the paper-recommended prompt is used automatically.
seedINT420–18446744073709550000Random seed. Paper default: 42.
stepsINT2812–100Number of denoising steps. Paper default: 28. The official demo allows 12-40. Lower = faster but less detail.
guidance_scaleFLOAT3.01–6CFG guidance scale. Paper default: 3.0. The official demo allows 1.0-6.0. Higher values follow the prompt more strongly.
size_levelINT1024256–4096Target resolution for processing. Paper default: 1024. The input image is resized so its total pixel area is roughly size_level x size_level (preserving aspect ratio), processed at that resolution, then resized back to original dimensions. Higher = better quality but more VRAM (~34GB at 1024).
device_strategyCOMBOautoHow to manage GPU memory. - auto: Uses full_gpu if >40GB VRAM, offload_to_cpu if >26GB, otherwise sequential_offload. - full_gpu: All components stay on GPU. Fastest. ~34GB at size 1024. - offload_to_cpu: Components move to CPU when not in use. Slower but ~24GB peak (the transformer weight size). - sequential_offload: Pre-loads as many transformer blocks to GPU as VRAM allows, streams the rest from CPU. On a 24GB card most blocks stay on GPU with only ~12 streaming. Enables cards that can't fit the full transformer.

Outputs (1)

NameTypeDescription
imageIMAGE