ComfyUI: Imposing Consistent Light (IC-Light Workflow Tutorial)

The video focuses on implementing IC-Light in Comfy UI, specifically for product photography. IC-Light is based on SD1.5, and we use a reference background and a product/object photo to regenerate the background and re-light the object. Images are generated in SDXL resolution, then upscaled by 4x. A number of unique techniques are used to transfer details to the final generation and even on the 4 K upscaled images.

------------------------

Links for Models:

GitHub:

ComfyUI Official:

------------------------

TimeStamps:

0:00 Intro.
01:14 Requirements.
06:50 Res Calculation, Masking.
14:21 Product Composition, Masking, Ollama.
23:15 ResAdapter, ICLight, Sampling, Detail Restore.
38:11 Workflow Test, Masking, CFG.
40:38 H/L Frequency Masking Blur, Directional Lighting, BG Processing.
52:23 Detail Transfer Upscaled Images.

[controlaltai]

Comfy Update (Aug 27, 2024): If you Are getting KSampler Error, You need to update ComfyUI to "ComfyUI: 2622[38c22e](2024-08-27)" or higher, IC-Light and Layered Diffusion. Everything works as shown in the video. No change in workflow.

IC-Light is based on SD1.5, but all generates are SDXL Resolution, then 4x Upscale. I hope you find the tutorial helpful. Please note: at 5:17 layered diffusion custom node is needed, even though no nodes are used otherwise you will get an error as follows:

RuntimeError: Given groups=1, weight of size [320, 4, 3, 3], expected input[2, 8, 64, 64] to have 4 channels, but got 8 channels instead

[esuvari]

Oh MY GOD! This is incredible! The first two random images I tried off the top turned out amazing, first try. You're the most underrated SD channel on youtube, thank you for this amazing work. Can't wait to get my hands dirty with this. Wish you the best.

[kobe]

honestly this is too good, thank you so much

[GoodArt]

that was just the coolest video I've ever seen. comfy rules.

[dankazama]

can we have this kind of workflow with flux? this video deserve more views. Good work sir/ma'am!

[jd]

Yes, finally! thank you for this tutorial

[PrithivThanga]

Must be a worthy one. will test and post here..

[oohlala]

Thank you for this tutorial. However, I don't understand why we need to segment the image again at 16:51. We already have the mask and the image with the new composition (product size and placement) as output of the "ImageBlendAdvance V2" node. Why are we repeating the segmentation process? The resulting images and masks of the new segmentation seem to me to be the same as the outputs from the "ImageBlendAdvance V2" . Sorry to ask about that. I'm a sub, and thoroughly enjoy your tutorials.

[LX]

How many times did you paused this video?
R: Yes :))

[DanielPartzsch]

Very nice. What exactly is the difference between the old IC light models and the once you've used here? Do they yield better results? Thanks

[ivanivan]

Hello, thank you for the course, I've realized the whole workflow, and I just can't make the product look transparent when placing transparent glass products such as perfume and wine glasses, it can't show the background content through the glass at all, I've watch the tutorials again and again and didn't find where to set the transparency of the product, may I ask where to set it? Looking forward to your answer, thanks!

[josephmorgans]

great work, thank you !
Is it possible to edit / change the background & product (STRING) promts ?

[ronshalev]

Is it possible to control the image background blur result?

[design]

Hi, great tutorial, by the way! I have a slight problem. The resulting image of a black product is different from the original. For example, if the product is black running shoes and the background is a green scenery, the result will make the shoes appear green. I also tried a black bag, and it turned white. the details is still there, but this result is after ksampler. probably has something to do with the IPAdapter or IClight ??

[SteMax-dz]

at upscale part, i ImapactInt = 2, the product image get bigger， it bigger than bg imgae. i dont know why, sir help

[SaoirseChen-vb]

Thank you for the incredible workflow, I got an issue when generating the ksampler before the details and color adjust parts, the ksampler image became total black at 60%, and the colormatch got error(stack expects a non-empty TensorList), do you have any clues?

[blackbear]

Hi, is there a way that i can make the background less cartoonish? I already try many checkpoint but it give the same result. How do i make realistic background. I already use realistic image for the background image though.

[Andrew-hilk]

This is amazing!

Any ideas about this error?
Error occurred when executing KSampler:
Given groups=1, weight of size [320, 4, 3, 3], expected input[1, 8, 104, 152] to have 4 channels, but got 8 channels instead

[sagarsinghvi]

Can you share the workflow for us to download?

[epfzoqx]

May I ask what caused this error

Error occurred when executing KSampler:
Given groups=1, weight of size [320, 4, 3, 3], expected input[2, 8, 128, 128] to have 4 channels, but got 8 channels instead