Nodes/ComfyUI-Rennart/🔍 Rennart Image Upscale (Model)
ComfyUI Node

🔍 Rennart Image Upscale (Model)

Load and upscale in one node — and finally get your hands on tile_size

By Rennart2025·Created 3 months ago·Updated 11 days ago· 4
🔍 Rennart Image Upscale (Model)
  • image
  • image
◄model_name▾►
◄tile_size512►
◄overlap32►
◄auto_reduce_on_oomtrue►

What this node is for

ComfyUI ships UpscaleModelLoader and ImageUpscaleWithModel, and the standard workflow is two nodes and a wire. Rennart Image Upscale (Model) collapses that pair into one - and then does what core refuses to do, which is let you set the tile size.

That's a good pitch, because tile_size is the first knob you want when a 4x pass dies on an 8GB card. If you've hit "Allocation on device 0 would exceed allowed memory" mid-run and had no lever to pull except "buy a bigger GPU", this node is the lever.

Be clear about which job this does. Upscaling is three unrelated jobs wearing one word: more pixels, more invented detail, more pixels over time. This is the first one - plain super-resolution in the ESRGAN family, loaded through spandrel, the same model-loading library ComfyUI itself uses. It makes the image larger; it cannot invent content and it will not rewrite a face. When your source is already sharp and just needs to be big, that "cannot" is a feature. When it's soft and you want detail pulled out of nothing, you want SeedVR2 or a tile-ControlNet pass, and no setting here gets you there.

One side benefit: tiled super-resolution doesn't suffer the classic tiled-diffusion failure where tiles drift from the source. A plain upscale model has no prompt to drift toward, so ControlNet Tile isn't needed here; overlap is about seams, not coherence.

How it works

The node loads the named model from ComfyUI/models/upscale_models, wraps it in ComfyUI's model patcher, and hands the image to comfy.utils.tiled_scale - which walks the image tile by tile, runs the model on each, and feathers the overlaps into a single full-size output. Output dimensions are whatever the model's scale factor dictates: a 4x model turns 1024 into 4096, with no separate factor widget to fight with.

Memory handling mirrors core's upscaler: it asks for enough VRAM to fully load the model (force_full_load), so on a tight card ComfyUI evicts your diffusion model rather than refusing. If a pass still throws, auto_reduce_on_oom halves the tile and overlap and retries, down to 128px, then re-raises. Progress is a real bar, counting batch × tiles.

An optional spandrel_extra_arches import runs at load: with that package installed, architectures core's spandrel build won't accept become loadable here.

The inputs you actually set

  • model_name - a dropdown built from the files in upscale_models. Subfolders show up as paths (4x/4x-UltraSharp.pth), so organising it is free. An empty dropdown means an empty folder.
  • image - an IMAGE batch. Hand it the whole batch in one call: the model loads inside the node every execution, so one call for eight frames is one load, eight calls are eight.
  • tile_size (64–4096, default 512) - the author's own tooltip says it plainly: bigger is faster but needs more VRAM. Peak memory scales roughly with tile², so 1024 tiles are around four times the pressure of 512 for maybe 1.5x the speed.
  • overlap (0–512, default 32) - the seam fuse. Watch out here: set overlap to half the tile or more and the node silently resets it to tile_size//4. Ask for 256 on 512 tiles and you get 128, with no warning.
  • auto_reduce_on_oom - leave it on unless you specifically want a loud failure.

One output, image, same format, larger. Wire it into SaveImage, or into a face pass if the upscale left the skin looking plasticky.

Installing it

ComfyUI Manager → search ComfyUI-Rennart → install → restart. Or by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/Rennart2025/ComfyUI-Rennart
# restart ComfyUI

No pip step: requirements.txt is torch and numpy. The models are your own problem - grab a 4x ESRGAN-family file (4x-UltraSharp, Remacri, NMKD-Siax) from OpenModelDB or Civitai and drop it in ComfyUI/models/upscale_models.

Troubleshooting

It's not in the node menu. The pack's __init__.py imports each file and swallows exceptions into a console warning, so a missing dependency quietly disables one node. Grep the log for [Rennart] - the text is Russian (Не удалось загрузить ..., "failed to load ...") and names the file that broke. This node needs spandrel, which ships with ComfyUI, so it's usually a stripped-down install.

Grid lines, or a blurry patchwork. Overlap too small for the tile size, often after the OOM loop quietly halved your tile on a multi-image batch. Try 512 tiles with 64 overlap and one image at a time. Related: small tiles mean many more tiles, and tiled inference is dominated by per-tile overhead, so reaching for 2048 tiles is often faster as well as cleaner - right up until VRAM says no.

Faces come out plasticky. Upscaling bakes in what was already there, compression damage included. Fix defects at native resolution first; if you wanted detail invented rather than preserved, this isn't the node for that job.

CategoryRennart/Upscale

Inputs (5)

NameTypeDefaultDescription
model_nameCOMBO0 options:
imageIMAGE—
tile_sizeINT51264–4096Размер тайла (плитки) в пикселях. Больше = быстрее, но требует больше VRAM.
overlapINT320–512Перекрытие между тайлами в пикселях. Помогает убрать видимые швы.
auto_reduce_on_oomBOOLEANtrueАвтоматически уменьшать размер тайла при нехватке VRAM.

Outputs (1)

NameTypeDescription
imageIMAGE—