Relight any photo with IC-Light in ComfyUI: prompts, light masks and detail transfer

IC-Light relights a photo from a text prompt or a light mask. It does more than move the light: it also does a really nice retouching job, close to professional level, and you can feed it any rough photo of yourself, your family or a landscape. It runs in ComfyUI on almost any gaming laptop, at home, for free, without uploading your photos anywhere.

IC-Light is by lllyasviel, the author of ControlNet (IC-Light on GitHub). There’s also an IC-Light v2, which gives better results, but when I made this it was only available as a paid, non-commercial service on its own website, not in ComfyUI. Here I use v1.

New to ComfyUI? Start with ComfyUI node basics. The workflow and my light masks are free on my Patreon.

Install

1. Manager → Custom Nodes Manager, search IC-Light. There are two packs, from Kijai and IC-Light Native by huchenlei. Both are useful; I install both, and use Kijai’s here.
2. Also install KJNodes (Kijai’s utility nodes). Restart and reload the page.
3. Manager → Model Manager, search IC-Light. There are three models; you only need one:
   • FC (foreground conditioned): the one I use.
   • FBC (foreground and background).
   • FCON: according to the documentation, slightly worse than FC.
4. After the download, just refresh the page (Ctrl+R); no restart needed.

Build the workflow

Start from the default workflow and switch the checkpoint to epiCPhotoGasm. Rename the prompts “Positive” and “Negative”, and colour them green and red.

1. Load And Apply IC-Light: the checkpoint’s model goes through it into the KSampler. It picks up the FC model automatically.
2. IC-Light Conditioning: connect the positive prompt, the negative prompt and the VAE; its positive and negative outputs go to the KSampler.
3. The foreground: Load Image with your photo → Image Resize (Essentials) → VAE Encode → the conditioning’s foreground input.

The conditioning node also has an opt_background input. I tried background images and they never looked good, so I stopped; describe the background in the prompt instead. It has an empty_latent output too, the same as an Empty Latent Image node. I keep a separate Empty Latent Image, because that input is where light masks go later.

Type a lighting in the positive prompt, such as daylight or foggy morning, and press Ctrl+Enter. Each run is slightly different; raise the batch count to see many variations. Some will be horrible, some fine to work with. Cherry-pick.

Compare, and keep the face with Detail Transfer

Add rgthree’s Image Comparer with the result as image A and the resized original as image B. Drag the slider and you’ll see it changes the face slightly. That’s the retouching side: it “improves” you. You’re not exactly the same afterwards, which is more often bad than good.

To bring the detail back, add Detail Transfer: the relit output into target, the original image into source. It works like a high-pass, unsharp-mask trick in Photoshop, and it can be too much; lower the blend factor to control it.

Fix the seed while you tune

With the seed on randomize, every change re-runs everything and gives a new image. Switching to fixed doesn’t freeze the current image either, because the next seed has already been set. What I do instead: set the seed to increment, starting from 1, so each image has a known number. Go back to the number you liked (say 3), run it, then switch to fixed. Now when you change the Detail Transfer settings, only that node re-runs, in a few seconds.

Light masks

Instead of an empty latent, give the KSampler a light mask: a black-and-white image where white is light. Load the mask, VAE Encode it, and connect it to the KSampler’s latent_image. I drew a set of masks (round, stripes, diagonal lines and others); they’re on my Patreon.

Hard masks burn the image. A pure black-and-white mask gives super high contrast, and a round mask produced a very strong light that also hit the eyes. To soften it, add Remap Image Range between the mask and VAE Encode, with a preview: raise the minimum and lower the maximum. I went down to about 20 percent. It’s not real-time like Photoshop; you run the graph to see even the preview.

With masks, the prompt matters a lot. “Spotlight” gave harsh light; “neon light” gave neon. I usually queue six or eight at a time. (Clicking the number on a preview node shows all the previews from earlier runs, a newer ComfyUI feature; update ComfyUI now and then.)

Multiplier: IC-Light Conditioning’s default 0.182 is good. At 0.155 it got dimmer; at 0.25 it burned the image almost into a stylised look.

Black-and-white results

Sometimes it returns black-and-white images and I don’t know why. Putting “black and white” in the negative prompt made the saturation go crazy. Trying the Photon model also produced plenty of black-and-white images at first. My 4 GB laptop made them much more often than my 24 GB PC, which mostly gives colour. A prompt like “50% saturated” can help. In the end I mostly cherry-picked rather than prompting a lot; I keep prompts minimal.

Higher resolution

SD 1.5 and IC-Light are trained on 512 × 512. You can still feed a 1024 × 1024 image, as long as the mask is the same size (use a resize node). The result was a bit burned and more stylised, but higher quality. That can be worth it, because upscalers sometimes soften skin in a plastic way.

Paint the light by hand

You can also paint a light mask directly:

1. Right-click Load Image → Open in Mask Editor, and paint where the light should fall (I tried a 1950s noir look on a face).
2. Convert Mask to Image, then into VAE Encode.
3. If the result is blurry, check the sizes: when the mask doesn’t match the 512 × 512 image exactly, it breaks everything.
4. Soften it: Mask Blur (Essentials) before converting to an image (I used 66), then Remap Image Range at about 30 percent with the whites lowered.

Groups and landscapes

IC-Light also works on several people. On a generated family photo, “daylight” gave interesting results. A beam-of-light mask on the child, or “beam of light” in the prompt, created some crazy light beams; the multiplier may need adjusting.

On a landscape photo from Australia, “daylight” gave interesting results, “sunset” made it far too red and scary, and “afternoon” sometimes replaced the sea with a landscape that wasn’t there, but sometimes did a good job. You have to play with the prompts.

Upscaling the result

A relit 512 px image can go to 8K with my free upscaling workflow. The result wasn’t always amazing; I often found the 512 relit image better. But if you need the resolution, it still gives a decent result.

Anton Tenitsky is a 3D artist with 15+ years in computer graphics, working with Nickelodeon, Gameloft, Marvel Rivals and many other clients. More about me