Image Detail Transfer

Ever run a face through an upscaler or a smoothing pass and end up with plastic, poreless skin? This node is the fix. easy imageDetailTransfer takes the fine, high-frequency detail from one image (the source) and lays it over another (the target), so you can put the texture back - skin pores, fabric weave, surface grain - without touching the underlying shapes and colors. It's the frequency-separation trick that retouchers have used for years, packaged as a single ComfyUI node.

How it works

The mechanism is old-school image compositing, not diffusion. The node blurs the source to isolate its low frequencies, subtracts that to get the high-frequency detail layer (the fine texture), and blends that detail onto the target using a chosen blend mode. Because it's only moving the fine stuff, the target keeps its overall structure and tone; it just gains crispness and texture from the source. That's why it pairs so well after any step that softens an image - upscaling, denoising, face restoration, IP-Adapter passes.

The inputs and outputs that matter

• target - the image receiving detail (your smoothed / upscaled result).
• source - the image the detail comes from (often the original, pre-smoothing version).
• blur_sigma - the frequency split. This is the knob that decides what counts as "detail." Small values grab only the very finest texture; larger values pull in coarser features. Start near the default and nudge until pores show without the whole image getting crunchy.
• blend_factor - how strongly the detail layer is applied. Turn it down if the effect looks harsh or over-sharpened; the range even allows negatives if you want to remove detail.
• mode - the blend mode (add, soft_light, overlay, multiply, and eight more). add is the default and the honest detail-transfer default; soft_light and overlay give a gentler, more contrast-aware result. Worth trying a couple on your image.

Optional mask restricts the transfer to a region - perfect for "put texture back on the face but leave the smooth background alone." The single output is the composited image, plus the usual image_output / save_prefix since this is an output node.

Installing it

Comes with the pack. Install ComfyUI-Easy-Use via ComfyUI Manager (search ComfyUI Easy Use, restart) or:

cd ComfyUI/custom_nodes
git clone https://github.com/yolain/ComfyUI-Easy-Use

Install requirements, restart. No models - this is straight image math, so it's fast and CPU-cheap.

Where it bites

The classic mistake is mismatched source and target. The two images need to line up - same subject, same framing, same resolution - or you'll be pasting detail onto the wrong pixels and it'll look like a ghosting mess. If your source is the pre-upscale original, resize it to match the target first.

The other is overdoing it. Crank blend_factor too high or pick too small a blur_sigma and you get that over-sharpened, gritty HDR look that screams "processed." Detail transfer is seasoning, not the main course - aim for "wait, is this even edited?" and stop there. Use the mask to keep the effect off areas that shouldn't have texture (skies, out-of-focus backgrounds), and you'll get clean, believable results.