HDR Grain Matcher

Grain is the cheapest realism upgrade there is, and the community has been saying it for years - as one commenter put it in a thread about making AI images read as photographs, "I also add grain and it does a lot for realism," right next to "motion blur + camera shake + grain + red channel aberration + vignetting." The catch is that grain added the obvious way - additive noise on a 0–1 image - looks like video noise, and it destroys highlights, because noise added in display space is enormous relative to a bright pixel's value and invisible on a dark one.

HDR Grain Matcher does it the way a film pipeline does. It extracts grain from a reference plate, works in log2 exposure space, and transfers it so the grain scales with exposure - dense in the shadows, subtle in the highlights - which is what real film does.

How it works

The reference plate is pushed into log2 space, box-blurred, and the blur is subtracted from the original. Whatever survives is high-frequency detail: grain, texture, sensor noise, all of it. That residual is then added to the target in log2, which is what makes the magnitude exposure-relative rather than absolute - adding a fixed amount in log space is a multiplicative change in linear light. Same trick as exposure itself.

The blur is F.avg_pool2d with your kernel_size, so it's a fixed box filter, not an edge-aware one. That has a consequence: anything high-frequency in the reference comes with it.

Then each channel gets its own multiplier (r_gain, g_gain, b_gain) so you can match the reference plate's colour grain - real stocks are not neutral.

Practically: kernel_size (odd, 1–15, default 3) is the main knob. Small captures fine grain and little else; large starts dragging coarser texture across. intensity is a straight multiplier on the extracted grain, where 1 means "matched to the reference." One output: grained_image.

Install

Manager → search Radiance → Install → restart → refresh your browser.

cd ComfyUI/custom_nodes
git clone https://github.com/fxtd-studios/radiance.git
cd radiance
python -m pip install -r requirements.txt

Windows portable users should use ComfyUI's bundled python_embeded\python.exe for the pip step. No models to download - this is a pure tensor operation.

Where people get burned

Pick a flat reference. Because the node transfers all high-frequency detail, a busy reference plate doesn't give you its grain, it gives you its content - foliage, fabric, skin texture - stamped onto your target. The tooltip says it outright: use a flat, defocused area if possible. This is the single biggest quality difference between a good and a bad result here.

The reference must match the target's width and height. No resizing happens for you. A 4K plate against a 1080p target will error or misbehave; the frames do cycle if the reference is shorter than the target batch, so a short clip is fine, a differently-sized one isn't.

kernel_size of 1 extracts nothing - the tooltip is honest that the box blur degenerates and you get a no-op. Use odd numbers; the widget steps by 2 anyway.

And the practical one on long shots: process a few test frames before queueing 240. Grain extraction is cheap, but you'll want to eyeball intensity against a real frame rather than a still.