Save 4 hours! We auto-setup your workflow! Free!

Drop your workflow.json — we handle every dependency, custom node, and model. Just open the link and run.

Auto-Setup Workflow Json (Free) Now!
ComfyUI > Nodes > ComfyUI-PhotoDeblur > Deblur Image (Tiled) · Photo Deblur

ComfyUI Node: Deblur Image (Tiled) · Photo Deblur

Class Name

PhotoDeblurTiled

Category
Photo Deblur
Author
porfeng (Account age: 756days)
Extension
ComfyUI-PhotoDeblur
Latest Updated
2026-11-11
Github Stars
0.09K

How to Install ComfyUI-PhotoDeblur

Install this extension via the ComfyUI Manager by searching for ComfyUI-PhotoDeblur
  • 1. Click the Manager button in the main menu
  • 2. Select Custom Nodes Manager button
  • 3. Enter ComfyUI-PhotoDeblur in the search bar
After installation, click the Restart button to restart ComfyUI. Then, manually refresh your browser to clear the cache and access the updated list of nodes.

Visit ComfyUI Online for ready-to-use ComfyUI environment

  • Free trial available
  • 16GB VRAM to 80GB VRAM GPU machines
  • 400+ preloaded models/nodes
  • Freedom to upload custom models/nodes
  • 200+ ready-to-run workflows
  • 100% private workspace with up to 200GB storage
  • Dedicated Support

Run ComfyUI Online

Deblur Image (Tiled) · Photo Deblur Description

Deblurs large photos using overlapping tiles, blending seams while preserving image size.

Deblur Image (Tiled) · Photo Deblur:

PhotoDeblurTiled restores blurred photographs by processing them in smaller, overlapping sections and combining those sections into a full-size image. This tiled approach makes large photographs more manageable because the original image and full-size working buffers stay in system memory while individual patches are sent to the model’s processing device. Overlapping regions are blended to reduce visible boundaries between restored sections. The node performs deblurring rather than upscaling, so it preserves the image’s dimensions, and it preserves existing transparency without applying restoration to the transparency channel. Optional test-time augmentation performs additional transformed evaluations, offering another restoration approach at substantially greater computational cost, although it can soften some details. The restoration behavior depends on the loaded model, and the node keeps your chosen tile settings rather than silently reducing them when GPU memory is insufficient.

Deblur Image (Tiled) · Photo Deblur Input Parameters:

image

The photograph or image batch you want to deblur, supplied as a required IMAGE connection. It must be a nonempty ComfyUI image arranged as batch, height, width, and channels, with either three RGB color channels or four RGBA channels. All pixel values must be finite, meaning they cannot contain invalid numerical values such as NaN or infinity. For an RGBA image, only the RGB channels are restored; the alpha channel, which controls transparency, is saved separately and reattached unchanged afterward. The image is copied into CPU memory as 32-bit floating-point data before tiling, helping avoid placing the entire photograph on the GPU at once. There is no numeric default, minimum, or maximum specified for this connection, but the batch size, height, and width must each be at least 1.

deblur_model

The restoration model used to process each image tile, supplied as a required PHOTO_DEBLUR_MODEL connection from a compatible Photo Deblur loader. This connection carries the model together with information about its family, variant, and processing device; it is not interchangeable with an ordinary diffusion MODEL connection. The selected model determines the underlying restoration method, while this node handles tiling and recombination. The provided context includes the HI-Diff loader, and the restoration code also explicitly handles the MISCFilter family. MISCFilter requires a CUDA device and will not run through this node on a non-CUDA device. The model is requested as a full load through ComfyUI’s memory management, so tiling reduces patch-related memory requirements but does not eliminate the memory needed for the model itself. There is no default value or numeric range for this connection.

tile_size

The side length, in pixels, of the square tiles used for restoration. The default is 768, the interface minimum is 64, the maximum is 4096, and the adjustment step is 8; the value must be a multiple of 8. Larger tiles generally reduce the number of patches needed to cover a photograph but require more processing memory per patch. Smaller tiles reduce patch memory requirements and can help when GPU memory is limited, although they generally require more tile evaluations for the same image. The node’s memory estimate increases with the square of the effective tile side, so changing this setting can have a substantial effect on memory demand. The overlap must remain smaller than this value. If GPU memory runs out during tiled processing, the node recommends reducing this setting to 512 rather than automatically changing it for you.

overlap

The width, in pixels, of the shared area between neighboring tiles. The default is 96, the interface minimum is 0, the maximum is 2048, and the adjustment step is 8. In addition to these interface limits, the value must be smaller than tile_size. Shared areas allow restored tiles to be blended instead of meeting abruptly at a hard boundary, which can help reduce visible seams. Increasing the overlap generally means that more image regions are processed repeatedly, increasing the amount of work and potentially the processing time. An overlap of 0 is accepted, but it provides no shared region for blending neighboring tiles. If you reduce tile_size, check this setting as well; for example, tile_size 512 with overlap 96 is valid, while tile_size 512 with overlap 512 is not.

blend

The method used to combine restored tiles into one image, supplied as a required string selection. The default is comfy_linear, which uses ComfyUI’s tiled blending implementation at an output scale of 1, preserving image dimensions. Other accepted selections are defined by the extension’s BLENDS collection and use its weighted-tile implementation. The provided source does not show the contents of that collection, so the complete list of available options cannot be established from this context; use the options actually displayed in your installed node rather than entering an unverified name. This setting affects how overlapping tile results are merged and can therefore influence transitions between neighboring patches. There is no numeric minimum or maximum, and any selection outside BLENDS is rejected.

tta

Enables or disables test-time augmentation, an optional restoration mode that uses eight transforms during prediction. The default is False, and the available values are True and False. When enabled, the node accounts for eight model evaluations per tile in its progress tracking, and the interface warns that this involves about eight times the model work. This may be worth comparing with ordinary processing for a particular photograph, but it is not a guaranteed quality improvement: the tooltip specifically notes that it can soften some details. Leave it disabled when you want a faster baseline result, and enable it only when the additional processing cost is acceptable.

seed

The integer starting seed passed to the tile prediction process. The default is 100, the minimum is 0, and the maximum is 18446744073709551615. The node passes the starting seed to the first prediction call and increases it by one for each subsequent tile prediction call, giving different calls successive seed values rather than reusing one seed everywhere. Its visible effect depends on whether the loaded model’s prediction process uses randomness. Keep it unchanged when comparing settings so that you are not also deliberately changing the starting random state, but do not assume that the seed alone guarantees identical results across different models, tile layouts, or execution environments. Automatic seed changes after generation are disabled in the parameter definition.

Deblur Image (Tiled) · Photo Deblur Output Parameters:

image

The restored photograph or image batch, returned as a single IMAGE output. It retains the original image dimensions and batch structure because the node performs restoration at a scale of 1 rather than enlarging the image. Restored RGB values are limited to the range from 0 to 1 before the output is returned, making the color result suitable for normal ComfyUI image workflows. If the original image contained an alpha channel, that channel is reattached unchanged after RGB processing; consequently, the color clamping does not alter the preserved alpha values. The full-size restoration result is assembled on the CPU and returned in ComfyUI’s batch, height, width, channels layout. You can connect it to an image preview, save node, or further image-processing nodes to inspect, export, or refine the restoration.

Deblur Image (Tiled) · Photo Deblur Usage Tips:

  • Start with the defaults: tile_size 768, overlap 96, blend comfy_linear, tta False, and seed 100. These give you a baseline before changing settings to address memory limits or visible artifacts.
  • If GPU memory runs out, try tile_size 512 while keeping overlap 96. If necessary, reduce the tile size further in multiples of 8, never below 64, and keep overlap smaller than the tile size.
  • If tile boundaries are visible, try increasing overlap in steps of 8 while staying below tile_size. For example, compare overlap 96 with 128 at tile_size 768, and check whether the improved transitions justify the extra processing.
  • Leave tta disabled for initial tests or large batches. Once you have a useful baseline, compare a result with tta enabled and inspect fine textures and edges, since the additional work can soften some details.
  • Keep the seed, model, and tiling settings fixed when checking repeatability. When evaluating a single setting, change only that setting so that differences are easier to interpret; changing tile size or overlap also changes how the image is divided into prediction calls.
  • Allow for both GPU memory and system RAM usage. Tiling keeps the full image off the GPU during patch processing, but the original image and full-size reconstruction buffers still occupy CPU memory; process fewer images at once if system memory becomes a problem.
  • Use MISCFilter only with a CUDA-capable setup configured for that model. Reducing tile_size cannot fix the explicit CUDA requirement.
  • If restoration introduces broad brightness or color shifts, you can optionally follow this node with PhotoDeblurToneMatch, connecting the original photograph and restored result. That separate node requires identical image dimensions and OpenCV; its default sigma is 16.0, and sigma 0 bypasses its correction.

Deblur Image (Tiled) · Photo Deblur Common Errors and Solutions:

Expected a nonempty ComfyUI RGB/RGBA IMAGE [batch,height,width,channels].

  • Explanation: The image connection does not contain the expected four-dimensional image batch, has a channel count other than three or four, or has an empty batch or image dimension. A mask or an incorrectly converted tensor can cause this problem.
  • Solution: Connect a standard ComfyUI IMAGE output containing RGB or RGBA data. Convert grayscale or other image formats into RGB before this node, and check that the image loader successfully produced a nonempty image.

tile_size must be a multiple of 8 >=64; overlap must be smaller than tile_size.

  • Explanation: This validation message appears if tile_size is below 64, is not divisible by 8, or overlap is negative or not smaller than tile_size. The same validation also rejects an unrecognized blend value, even though the message does not mention blending.
  • Solution: Choose a tile size such as 512 or 768 and an overlap such as 96. Ensure overlap is at least 0 and strictly smaller than tile_size, and select a blend option from the installed node’s dropdown. If workflow settings were imported or edited manually, reset blend to comfy_linear.

Input contains non-finite pixels.

  • Explanation: At least one input pixel contains NaN or infinity rather than a valid finite number. This usually indicates a problem in an earlier image-processing step rather than a tile-size issue.
  • Solution: Preview the image immediately before this node and trace back through earlier processing steps. Reload the original photograph or bypass the step producing invalid values, then supply a valid image again; changing the seed or overlap will not repair invalid input pixels.

MISCFilter requires CUDA.

  • Explanation: The supplied model belongs to the MISCFilter family, but its configured processing device is not a CUDA device. This family has an explicit CUDA requirement in the node.
  • Solution: Run MISCFilter in a working CUDA-capable ComfyUI environment and ensure the model is configured to use that device. If CUDA is unavailable, choose a different model only after checking that model’s device requirements; smaller tiles do not remove this restriction.

PhotoDeblur ran out of GPU memory. Reduce tile_size to 512; the node keeps your chosen settings and does not silently shrink tiles.

  • Explanation: A CUDA out-of-memory error occurred during tiled restoration. The current model and tile processing exceeded available GPU memory, and the node intentionally stopped rather than silently reducing your chosen tile size.
  • Solution: Reduce tile_size to 512, keep overlap below that value, and run again. If memory remains insufficient, try smaller valid tile sizes and close other GPU-heavy applications. The model itself also needs memory because the node requests a full model load.

GPU out-of-memory error during model loading; exact wording may vary.

  • Explanation: ComfyUI loads the model before entering the block that replaces tiled-processing memory errors with the node’s custom message. A failure at this earlier stage may therefore appear as a standard ComfyUI or PyTorch memory error instead.
  • Solution: Free GPU memory, close competing workloads, and try a smaller tile_size to lower the node’s estimated working-memory requirement. If the model itself cannot fit, smaller tiles may not be enough; consider a compatible model with lower memory requirements or hardware with more available GPU memory.

Deblur Image (Tiled) · Photo Deblur Related Nodes

Go back to the extension to check out more related nodes.
ComfyUI-PhotoDeblur
RunComfy
Copyright 2025 RunComfy. All Rights Reserved.

RunComfy is the premier ComfyUI platform, offering ComfyUI online environment and services, along with ComfyUI workflows featuring stunning visuals. RunComfy also provides AI Models, enabling artists to harness the latest AI tools to create incredible art.

Deblur Image (Tiled) · Photo Deblur