ComfyUI workflow
Baioni claims that native frequency separation in ComfyUI, built from Powered_JJ's node group and chained with background generation, subject masking and IC-Light relighting, restores the real product's fine details (such as printed text) after relighting and re-backgrounding, all with one click. In his words, this removes the two main obstacles to using generative AI in product photography: incoherent lighting and a product that is not the actual product.
The video names the commercial problem: in product photography you are 'not selling a 90% replica' but the actual product. A month earlier, generated shots were good enough but unconvincing, because the lighting did not match and the subject looked copy-pasted onto the background.
The breakthrough is presented as a community collaboration. A viewer challenged Baioni on a live stream to keep a whiskey bottle's fine details. Baioni first proposed Photoshop frequency separation, then Reddit user Powered_JJ posted a native ComfyUI version over the weekend. Together they integrated it into Baioni's existing relight workflow.
The workflow's logic is a staged progression, with each stage fixing the weakness of the one before. The original shot becomes a new background, then a blended composite, then a relit image. Because relighting degrades detail, original detail is restored last. Baioni says the video shows this progression is organic, and he repeatedly stresses that the demos are not cherry-picked.
The video states that one-click output is only a default. Preview-bridge masks, loaded masks and hand-drawn masks give the operator more control. The trade is that light directed this way is less organic than light taken from the generated background, but it is 'actionable' by the user.
The video states its own limits. It works at 1024x1024 because IC-Light only works with SD 1.5. Segmentation can miss parts of complex subjects. The frequency-separation math is tuned for average images and may need adjustment. People are not yet supported, because harsh lighting leaves shadow and light information in the detail layers.
Baioni claims the result may be more practically useful than LoRAs. LoRAs solve problems but are not one-click, whereas this is a one-click answer to location or complex studio shoots. He tells photographers they can either worry or get to know the technology and use it.
IC-Light relighting — A Stable Diffusion 1.5 relighting tool that relights a picture, here using white areas of the generated image as light sources via a mask. Apply: Feed the blended subject-plus-background image and the light mask into the relight group, then give it a relight prompt such as 'white light', 'neon light' or 'moonlight'.
Frequency separation — An approximation of the Photoshop technique that splits an image into a high-frequency layer (details) and a low-frequency layer (color and lighting), applied to both the original and the relit image. Apply: Run the original and relit images through the frequency-separation groups in the optional 'Preserve Details Such as Words from the Original Image' group to get their separate high- and low-frequency layers.
Mask-based high-frequency blend — A blend of the original image's high-frequency layer and the relit image's high-frequency layer, so the result has the original subject's details everywhere except the generated parts. Apply: Feed both high-frequency layers and the Segment Anything mask into an image-blend-by-mask node, then merge the blended layer on top of the relit image's low-frequency layer.
Gaussian blur radius tuning — The main adjustable parameter of the frequency-separation math; the default is set for the average use case. Apply: Lower the image Gaussian blur radius to keep finer details and raise it to keep fewer. Set the same values in every frequency-separation group or it won't work.
Depth and lineart background generation — A background generator that creates a new background from a prompt, guided by the depth and lineart of the original image. Apply: Write a background prompt and keep the lineart control for transparent subjects, or disable it if you don't want it.
Segment Anything with GroundingDINO subject compositing — A segmentation group that uses a text noun for the subject (bottle, bag, microphone) to mask the original subject, then merges it on top of the generated background. The mask is reused later in the workflow. Apply: Change the noun in the GroundingDINO node to match the subject in each new image, for example from 'bottle' to 'bag' to 'microphone'.
White-spot light-source mask — The default automatic way of defining light, taking the white areas of the generated image as a mask of light sources, passed through a grow-mask-with-blur node. Apply: Leave the default connections in place for a one-click, organic-looking result.
Hand-drawn lighting mask via Preview Bridge — A manual override where you draw the lighting mask yourself in the mask editor of a Preview Bridge node instead of sourcing it from the background's white areas. Apply: Open the mask editor, draw the light, then swap the mask-from-color output for the Preview Bridge mask and connect it to the grow mask with blur input.
Loaded lighting-shape mask — A bypassed Load Image Mask node that lets you use a prepared mask as the light source, such as strip lights coming from the sides. Apply: Enable the Load Image Mask node, load your mask, and link it to the grow mask with blur node.
Hand-drawn detail-preservation mask — An override that replaces the automatic Segment Anything mask with a mask you paint over only the details you want preserved, such as the inputs and the texture of the microphone head. Apply: Draw on the details in the mask editor, let the mask-preparation nodes process it, and connect the convert-mask-to-image output to the image blend by mask input in place of the Segment Anything mask.
Upscaling placement — Guidance on where to add an upscaler to a workflow fixed at 1024x1024. Apply: Add the upscaling group between the relight group and the detail-preservation (frequency separation) group, and resize the Segment Anything mask to match.
Multiple Segment Anything groups — A way to handle complex scenes where one segmentation pass can't capture the whole subject, such as the microphone with its arm. Apply: Use more than one Segment Anything group for complex scenes, and tinker with the prompt. For clear scenes with one or two subjects, one group is enough.
Generate-from-scratch product group — An optional group, currently not working, that generates a product from nothing. Apply: Use it to test the pipeline when you have no shot to work with; it isn't needed when you start from a real product photo.
One-click prompt-swap routine — The per-image procedure for reusing the workflow on a new product. Apply: Change the background prompt, copy it into the relight group's positive prompt and adjust the light descriptor (for example 'white light' to 'neon light'), change the GroundingDINO subject noun, then Queue prompt.
Frequency separation lets the workflow source lighting and detail independently. The generated image supplies color and lighting, and the original supplies fine detail.
The segmentation mask does double duty. It first isolates the subject for compositing, then it decides where original high-frequency detail is used (the subject) and where generated detail is kept (everything else).
Adding lineart control alongside depth is a targeted fix for transparent objects, where depth alone was not enough in Baioni's experience. It can be disabled.
If you add upscaling, place it between the relight group and the detail-preservation group, and resize the segmentation mask too. Otherwise the high-frequency blend is masked at 1024x1024.
The Gaussian blur radius is the main tuning knob: lower it to keep finer details, raise it to keep fewer. The value has to be identical in every frequency-separation group or the technique fails.
The microphone demo shows an incidental limit. GroundingDINO selected the microphone but not its supporting arm, so the arm was changed along with the background, and getting the whole object into the mask may need prompt tinkering.
People are not yet supported because, under harsh lighting, the high-frequency layers retain a lot of shadow and light information. That would carry the original lighting into the final image.
Baioni notes that workflows have become so complex that a from-scratch explanation would run about 40 minutes. He points viewers to a roughly 2-hour live stream for his train of thought, and says the many notes in the workflow are meant to prevent misuse, such as loading the wrong model in the IC-Light group.
«invaluable, of course, in product photography, where you're not selling a 90% replica but the actual product you are trying to sell.»
— 02:47
«The ability to have coherent lighting and the ability to have the actual product in your shots were some of the major obstacles up until now in generative AI in production environments, and during this weekend, we blew them out of the water.»
— 02:55
«IC-Light, which is a fantastic piece of tech — I'm not even kidding; I think it's the best thing to come to generative AI since LORAs»
— 05:17
«the details have been preserved, the light has been changed, and everything is working great.»
— 09:31
«So product photographers, not everything's over yet.»
— 16:33
«for harsh lighting conditions, the high-frequency layers keep a lot of shadow and light information in them»
— 17:06
«this is a one-click solution to location shoots or complex studio shoots.»
— 18:15
«If I were a photographer —and I am— I would be really worried about this.»
— 18:15
Reception
Viewers are enthusiastic about the workflow's results and value, but the reception is tempered by many setup difficulties and a recurring 'Image Levels Adjustment' math domain error that the author hasn't had time to fix, with the community sharing workarounds.
A dense, demo-driven walkthrough whose central contribution is restoring original product detail after relighting by blending frequency-separated layers under a segmentation mask. It is openly enthusiastic, and it states its own limits: 1024px SD 1.5, segmentation gaps, math tuned for average images, and no support yet for people.

19:02