Nodes/ComfyUI-Rennart/🧹 Rennart Denoise
ComfyUI Node

🧹 Rennart Denoise

Rennart Denoise runs your image through the model twice and averages the results. That's the good idea.

By Rennart2025·Created 3 months ago·Updated 11 days ago· 4
🧹 Rennart Denoise
  • image
  • image
  • delicate
  • aggressive
  • info
  • stat
◄strength100►
◄delicate_aggressive25►
◄delicate_gan_psnr25►
◄aggressive_plksr_esrgan25►
◄esrgan_inverted_direct50►
◄deblurtrue►
◄esrgan_downscaleLanczos►
◄tile_size512►
◄precisionfp32►

What it actually does

This is a cleanup pass, not an upscaler - 1x, same output size, minus the noise.

The model is 1xDeNoise_realplksr_otf: a RealPLKSR network, the fast lightweight architecture that landed in the community's upscaler talk back in 2024, trained on on-the-fly degradation - random noise and JPEG levels generated during training rather than one fixed set of damaged images. The file is small, tens of megabytes rather than gigabytes, and the node fetches it from Hugging Face on first run.

What you reach for it for: compression-krangled web images, phone shots in bad light, scans, old JPEGs with that smeared block texture. It repairs damage without generating, which makes it the cheap deterministic cousin of SUPIR - and the one difference that matters is that it cannot invent a new face. Point a denoiser at a portrait and you get the same portrait, cleaner. Point SUPIR at a portrait and you may get a plausible stranger.

Repair at native resolution first, then upscale: feed a noisy image to a 4x model and you've baked the noise in at four times the size.

The two-pass trick, which is the interesting part

Tiled inference is how this fits on modest VRAM, and it leaves seams: each tile is seen with hard edges, so the boundary can show as a faint grid. Every tiled upscaler has some version of this problem.

Rennart Denoise's answer is cheap and clever. It runs the image through the model once. Then it torch.rolls the image by an offset in X and Y - a cyclic shift, so pixels wrap around instead of leaving blank bands - runs the model again, rolls the result back, and averages the two outputs 50/50.

The tile grid now lands in different places, relative to the picture, on the two passes. Wherever the first pass left a faint boundary artefact, the second pass has a different boundary position, and averaging halves the visible error. Burst-stacking logic, applied to tiling - and the cost is honest: everything runs twice, so it's twice as slow as a single-pass denoise. Mechanically it's ComfyUI's own tiled_scale at 1x, in a model patcher, result clamped to 0–1.

Inputs: mode is basically the whole interface

  • mode - three choices, and this trips people up. auto ignores your tile_size, overlap and offset widgets entirely, using 512 / 128 / 192. If you're changing tile_size in auto mode and nothing happens, that's why. tile_size mode is the useful one: you set the tile and the rest is derived (overlap = tile/4, offset = (tile − overlap)/2). manual uses all three as typed.
  • tile_size (64–4096, default 512) - the VRAM dial. Bigger tile, fewer seams, more memory.
  • overlap (0–512, default 128) - how much neighbouring tiles share. Much more generous than the upscale node's 32, which makes sense: there's no size increase here to hide the artefact behind.
  • offset (0–1024, default 192) - how far the second pass is shifted. Around half a tile is the sane range; 0 defeats the whole point of the double pass.

Output is a single image, same dimensions as the input. There's no strength knob: the model runs at full strength, and the only blending is the internal 50/50. Too smooth? Blend the original back over it in a separate blend node. Batching works - send eight frames in one call rather than eight, since one execution means one model load.

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

Nothing to pip install - requirements.txt is torch and numpy. The model is handled for you: the first run downloads 1xDeNoise_realplksr_otf.safetensors from the Rennart/1xDeNoise_realplksr_otf Hugging Face repo into your first upscale_models path. On an offline machine, grab that file yourself and drop it in ComfyUI/models/upscale_models.

Troubleshooting

The node isn't in the menu. The pack imports its files one at a time and catches errors into a console print, so a broken file disappears quietly. Search your startup log for [Rennart] - the messages are in Russian, and Не удалось загрузить Rennart_Denoise ("failed to load") means the imports failed. This node needs spandrel and huggingface_hub, both of which normally ship with ComfyUI.

First run fails on the download. No internet, or the repo isn't reachable. Install the safetensors manually - and note the node only looks in the first upscale_models path, so if you've configured several, use that one.

Anime or flat line art comes out slightly sanded down. The model was trained on photographic degradation, so it treats crisp linework the way it treats grain. Test one image before running a folder through it; if that's your whole library, an edge-preserving filter is closer to what you want.

CategoryRennart/Image

Inputs (10)

NameTypeDefaultDescription
imageIMAGE—
strengthINT1000–100Общая сила шумоподавления: смешивание результата с исходником. 0 — исходник, 100 — полный результат.
delicate_aggressiveINT250–100Delicate ⟷ Aggressive: доля Aggressive (RealPLKSR + RealESRGAN) поверх Delicate (SCUNet). 0 — Aggressive не считается, 100 — Delicate не считается.
delicate_gan_psnrINT250–100Delicate: SCUNet GAN ⟷ SCUNet PSNR — доля PSNR поверх GAN.
aggressive_plksr_esrganINT250–100Aggressive: RealPLKSR ⟷ RealESRGAN — доля RealESRGAN (x2 и обратно) поверх RealPLKSR.
esrgan_inverted_directINT500–100RealESRGAN: инверсия ⟷ прямая — доля прямого прохода поверх прохода по инвертированной картинке.
deblurBOOLEANtrueDeBlur для обеих веток SCUNet (Delicate): резкость только по яркости (1x_DeBLR), цвет от SCUNet.
esrgan_downscaleCOMBOLanczosЧем вернуть выход RealESRGAN x2 к исходному размеру (Pillow). В графе — Lanczos.
tile_sizeINT512128–2048Тайл двойного прохода Rennart. overlap = tile/4, offset = (tile − overlap)/2. 512 — как в графе (overlap 128, offset 192).
precisionCOMBOfp32fp32 — точно как в графе. fp16 — быстрее и меньше памяти, результат может чуть отличаться.

Outputs (5)

NameTypeDescription
imageIMAGE—
delicateIMAGE—
aggressiveIMAGE—
infoSTRING—
statSTRING—