15:22
AI photo retouching
The video demonstrates a Freepik Spaces node-workflow that chains a ChatGPT-family "assistant" node (which analyzes an uploaded photo and writes a professional-retoucher-style prompt) into an image-generation node (Nano Banana / Nano Banana Pro) to fully automate high-end photo retouching with no manual sliders, and it teaches viewers to build, cost-optimize, and extend that same node pattern themselves.
A Freepik Spaces workflow uses ChatGPT (via an 'assistant' node) to analyze an uploaded photo, write a detailed retouching prompt like a professional retoucher would, and feed that prompt plus the image into an image-generation model (Nano Banana / Nano Banana Pro) — fully automatic, no manual sliders.
The node graph is: Media (input image) → Assistant (text model, analyzes image + writes prompt) → Image Generator (Nano Banana/Nano Banana Pro, takes prompt + reference image) → output.
Clicking a node's play button runs only that node (e.g., just generate the prompt); 'run from here' executes the whole downstream chain.
Generation settings (resolution, number of generations, model tier) directly drive credit cost; 4K plus multiple generations on the most intensive models can consume hundreds of credits (451 in the first demo).
Cropping the source photo to only the region needing retouching — instead of feeding the full high-resolution photo — lowers the effective resolution required, which helps hit 'unlimited' tiers and saves credits; the cropped, retouched result can then be re-aligned onto the full-res original in Photoshop.
Cropping to a square 1:1 aspect ratio (matching Nano Banana's preferred input) also makes results more consistent.
Setting resolution to 2K, toggling 'unlimited' on (for unlimited-plan users), and switching the assistant's model from a heavier one (GPT-5.2) to a lighter one (GPT-4.1 mini / GPT-5 Mini) can make a run consume zero credits.
To merge a retouched crop back into a full-resolution file: bring it in as a smart object, switch its blend mode to Difference, use Free Transform with an Alt/Option-clicked anchor point to precisely realign it over the original, then switch blend mode back to Normal.
Photoshop's Edit > Auto-Align Layers is offered as a faster alternative alignment method, but it requires rasterizing the layers, which sacrifices the smart object.
After alignment, Alt/Option-clicking the mask button creates a negative (black) mask; painting with a soft white brush selectively reveals only the AI-retouched areas, blending the result naturally with the original.
The same media→assistant→generator node pattern is reused to build a separate 'Makeup Space' that applies a described makeup look (or a generic 'high-end retouching' style) to a photo.
The makeup node only performs what it is explicitly instructed to do — run on an unretouched, blemished photo, it applies makeup but leaves blemishes untouched unless retouching instructions are also given; makeup and retouching outputs can be chained in either order.
Stated limitation 1 — resolution ceiling: Nano Banana tops out at 4K, and unlimited plans cap at 2K, which forces the crop-and-recombine workaround and becomes harder on very large images (e.g., full-body retouching on a 50-megapixel photo).
Stated limitation 2 — credit cost: multiple 4K generations on the most intensive models consume a lot of credits, mitigated by the unlimited-plan + 2K + lighter-model combination.
Stated limitation 3 — pixel regeneration: because the AI generates an entirely new set of pixels rather than editing existing ones, details can subtly change (e.g., an eye no longer looks like the same eye), which the video addresses by masking the original detail back in with a black brush on the layer mask.
Compared to the presenter's own 'Retouch for Me' plugin (skin tone, eye contrast, eye clean-up, depth, dodge and burn, all without generating new pixels), the AI and the plugin look similar unless you zoom in, but the AI approach handles extreme cases and complex makeup that the automatic plugin fails at or does incompletely — at the cost of altering some original pixels, which can be masked back.
Editorial framing: the presenter says they will keep using resolution- and pixel-preserving methods for primary work, treating this AI system as best for quick edits, 'impossible' fixes, or trying a different creative direction, and argues that professional value now lies in judging what result is great versus mediocre, not in operating the AI.
Freepik Spaces — A node-based AI workflow builder within Freepik where media, assistant, and generator nodes are wired together into a runnable pipeline. Apply: Go to freepik.com, click Spaces, and either open the presenter's pre-built retouching space via the link in the description or click 'new space' to build your own input-to-output pipeline.
Media node — The Freepik Spaces node that holds the input image (or other media) driving the workflow. Apply: Right-click in a space and choose Media, drop in a placeholder image, then click 'replace' and 'use image' to swap in your actual source photo.
Assistant node (ChatGPT-family text model) — A text-generation node in Freepik Spaces, backed by a ChatGPT-family model, that analyzes a linked image and writes a natural-language prompt such as detailed retouching instructions. Apply: Right-click and add an Assistant node, link the source image into it, type an instruction like 'analyze this image and write a detailed prompt to do high-end retouching,' then link its text output into the image generator node.
Image Generator node (Nano Banana / Nano Banana Pro) — The Freepik Spaces node that performs the actual image edit, using Google's Nano Banana or Nano Banana Pro model driven by the assistant's prompt plus the original image as a reference. Apply: Add an Image Generator node, select Nano Banana Pro (or Nano Banana 2) as the model, link in both the generated prompt and the reference image, and set resolution, generation count, and aspect ratio in its settings.
Assistant model swap for cost control (GPT-5.2 vs. GPT-4.1 mini / GPT-5 Mini) — Choosing a lighter text model for the assistant node changes its credit cost without necessarily changing the quality of the written prompt. Apply: In the assistant node's model dropdown, pick GPT-4.1 mini (or GPT-5 Mini) instead of GPT-5.2 so the prompt-writing step consumes no credits.
Crop-to-square (1:1) technique — Cropping the source photo down to only the region needing retouching, at a 1:1 aspect ratio, before feeding it into the workflow. Apply: In Photoshop or Freepik's built-in image editor, crop the photo to a square containing only the area to retouch, lowering the effective resolution needed, improving Nano Banana's output consistency, and enabling later re-alignment onto the full-res original.
Unlimited-plan + 2K resolution trick — A settings combination (2K output resolution plus the 'unlimited' toggle on an unlimited Freepik plan plus a lighter assistant model) that lets Nano Banana generations run at zero marginal credit cost. Apply: In the image generator node's settings, set resolution to 2K, toggle 'unlimited' on (requires an unlimited plan), set the desired generation count, and pair it with GPT-4.1 mini in the assistant node.
Difference blend-mode alignment — A Photoshop technique for precisely realigning a lower-res AI-generated crop over the original full-resolution photo as a smart object, using the Difference blend mode as visual alignment feedback. Apply: Drag the AI result onto the original as a smart object, set its blend mode to Difference, use Free Transform (Ctrl/Cmd+T) with an Alt/Option-clicked anchor/reference point to scale and move it until misaligned areas go black, then switch the blend mode back to Normal.
Edit > Auto-Align Layers — Photoshop's built-in automatic layer-alignment command, offered as a faster but less flexible alternative to the manual Difference-blend method since it requires rasterizing and loses the smart object. Apply: Duplicate the background layer, rasterize the AI-result layer, select both layers, then run Edit > Auto-Align Layers with 'Automatic' settings and click OK.
Negative layer mask + white-brush reveal — A masking technique to blend an AI-retouched smart-object layer with the original only in chosen areas, keeping the rest of the original intact. Apply: With the aligned smart-object layer selected, Alt/Option-click the layer mask button to add a black (hiding) mask, then paint over specific areas with a soft, white-foreground brush to reveal just the retouched regions you want.
Masking back original detail (black brush) — The inverse masking move used to correct places where AI-regenerated pixels changed something undesirably, such as an eye's appearance, by restoring the original detail. Apply: On the retouched layer's mask, paint with a black brush over the specific problem area (e.g., the eye) so the original pixels show through again.
Makeup Space — A second Freepik Spaces workflow, built on the same media→assistant→generator node pattern, dedicated to applying a described makeup look to a photo. Apply: Open the makeup space (in the same shared retouching space), replace the input image, type a makeup instruction such as 'dark red lipstick, and maybe slight blush,' then click the play button and 'run from here.'
Retouch for Me (with Dodge & Burn) — The presenter's own Photoshop retouching plugin, used as a non-generative comparison baseline that performs skin tone correction, eye contrast, eye clean-up, depth, and dodge-and-burn without generating new pixels. Apply: Run Filter > Retouch for Me > Heal (and its dodge-and-burn step) on a photo to produce an automatic, pixel-preserving retouch to compare against the AI-generated result.
The 'zero credits' result is not about doing less work — it is a specific configuration (crop to lower effective resolution + 2K output + the unlimited toggle + a lighter assistant model) that turns unlimited-plan generations effectively free.
Cropping to a square before generation solves three separate problems simultaneously: credit cost, the resolution ceiling, and output consistency, since Nano Banana behaves more predictably on 1:1 inputs.
Because the model regenerates pixels rather than editing them, the failure mode isn't 'looks bad' but 'looks good yet subtly wrong' (e.g., a changed eye) — which is why the recommended safeguard is localized masking of specific regions rather than rejecting or redoing the whole generation.
The three-node pattern (media → assistant writes a prompt → image generator executes it) is presented as a general-purpose template, not a retouching-specific pipeline — the video explicitly reuses it for a makeup space and implies it extends to old-photo restoration or 'any kind of retouching.'
The makeup node's behavior of only doing what it's told (not inferring unstated goals like blemish removal) means chaining order matters: retouch-then-makeup and makeup-then-retouch produce different results, giving users compositional control over the automated pipeline.
«What if you could just upload a photo and the AI automatically analyzes it, thinks like a professional retoucher, and sends precise instructions that fixes everything for you completely automatically?»
— 00:00
«It uses ChatGPT to analyze the image, think like a professional retoucher, and writes detailed instructions for high-end magazine-quality retouching while keeping the original skin texture and all the natural details intact.»
— 00:24
«Now this gives us unlimited literally unlimited.»
— 03:21
«Oh my gosh, all of the results are pretty impressive.»
— 10:47
«The next limitation with any AI is that it generates a completely new set of pixels, which is something you may or may not want.»
— 12:44
«Tools like this are insanely powerful, but they work best when they work alongside your existing workflow, not as a complete replacement.»
— 14:24
«What separates you as a professional from someone who just knows AI is that you know what is great and what is mediocre.»
— 14:30
Reception
Audience sharply divided between enthusiasm for AI efficiency and deep concerns about job displacement, skill devaluation, high costs, and ethical implications.
A practical, step-by-step tool demo with genuine before/after evidence and disclosed limitations (resolution caps, credit costs, pixel regeneration), while also functioning as promotion for Freepik Spaces and, via the comparison segment, for the presenter's own Retouch for Me plugin.
15:22