Photoshop 2027 beats GPT and Nano Banana in 3 of 6 tasks: cleanup, text, upscaling
Adobe has released Photoshop 2027 with 21 new features, and the author of the NeuroGraph channel compared the six main ones with GPT Image and Nano Banana Pro. To change the lighting on a portrait, a model needs only a phrase like "warm backlight from the left," while Photoshop charges 30 credits per attempt. But Photoshop does better at removing an object or noise without redrawing the whole frame. It won 3 of 6 comparisons, the models won 2, and one was a draw. Look for ideas in the models, then assemble the final image in the editor.
AI-processed from TG: @neyr0graph (NeuroGraph); edited by Hamidun News
Adobe has released Photoshop 2027 with 21 new features. Many of them do what generative models like GPT Image and Nano Banana Pro can already do from a single text prompt. For example, to reposition the light on a portrait, you just type "add a warm backlight from the left" in a model.
In Photoshop, the same task (the Relight feature) costs 30 credits per attempt. Credits are Adobe's internal currency for generation, and they are deducted even if you don't like the result. The author of the NeuroGraph channel went through the new features and compared six key ones with the models.
The outcome: Photoshop wins three of six, the models win two, and one is a draw. For designers, product managers and anyone building graphics pipelines, this hints at where to work in a model, where to work in the editor, and how to avoid wasting credits.
Where the models do just as well and Adobe asks for credits
Relight. The headline feature of the release: up to three light sources on a separate layer, with settings you can change later. But fine hair gets smeared, halos appear, and a glass bottle turns opaque. The models understand a prompt about lighting and usually keep the face intact. A model is enough for a sketch: Photoshop is easier to control, but the quality is still raw.
Generative Restyle (day to night, style changes). It looks impressive, but the intermediate frame costs 80 credits and loses quality. Both models can do the same thing. What's more, Restyle itself runs on a partner model such as Nano Banana. In effect, you are being sold the same model, only for Adobe credits. The models win.
Generative Recompose and Camera Angles (a different aspect ratio and a new camera angle). They cost from 10 to 90 credits. From a distance the result looks good, but up close the person has changed and the pattern on the shoes is different. Nano Banana Pro also changes both the aspect ratio and the angle, and it swaps out details in the same way. A draw. If you just need to extend the frame, the older Generative Expand is more reliable: it fills in the edges and leaves the original untouched.
Where Photoshop has no competition
Text editing. Imagine a flat JPEG with a banner where one word needs to change. Photoshop converts the text in the image into a real text layer with a similar font. This is a beta and costs about 10 credits. GPT Image will also replace the text on request, but it redraws it as part of the picture. A layer can be edited as many times as you like, so Photoshop wins.
Upscaling. Firefly Upscaler 2 scales an image up to 67 megapixels for 10–20 credits. It makes mistakes on small details: the stripes on a tie and the buttons drift. The models top out at 4K. Photoshop is ahead, but you need to check the result at 100%.
Cleanup and removal. Noise reduction now works not only with RAW but also with JPEG and PNG, and it doesn't invent anything. Remove deletes objects right on your computer, with no internet connection and no credits. The author calls it the best feature in the release. The models perform every edit by redrawing the entire frame, so Photoshop wins here outright.
What looks odd and dated
The author separately flagged what feels outdated:
- The built-in timelapse records only the canvas, and a regular screen recording captures more.
- Mockups: on a bottle, the design tears at the edges, and the old smart object with cylindrical warp handles it better.
- When asked to do frequency separation, the AI Assistant thinks for a long time and does nothing.
- Payment for every attempt: if you don't like the result, you pay again.
What to do with this
The pattern is simple. Where you need to invent a new image (light, style, camera angle), the plain models hold their own, and Adobe charges credits for the same job. Where you need to preserve the source (noise, removal, text in layers, large files), Photoshop has no competition.
The author's practical workflow: find the idea and sketch it in GPT Image or Nano Banana Pro, then assemble the final image in Photoshop with tools that protect your pixels. For teams, it is also a budget question: it makes more sense to run expensive lighting and style experiments in the models, and to pay Adobe only for what the models don't do.
Caveat: this is one author's comparison on his own examples, not a lab test. Lighting, style and camera angle depend heavily on the specific frame, so results on your files may differ.
- Nov 18, 2026KandiShop raises AI face similarity to the original from 52% to 75% without fine-tuning
- Nov 18, 2026Hugging Face routes LLM requests through OpenRouter, cutting capacity errors by 46%
- Nov 18, 2026CogAlign teaches Qwen3-VL-4B to pick the odd one out like humans: 53% → 57% agreement
- Nov 18, 2026Dreamina Launches Motion Studio: Motion Graphics Written as Code From a Prompt
Key Questions Answered
Where the models do just as well and Adobe asks for credits?
Relight. The headline feature of the release: up to three light sources on a separate layer, with settings you can change later. But fine hair gets smeared, halos appear, and a glass bottle turns opaque. The models understand a prompt about lighting and usually keep the face intact. A model is enough for a sketch: Photoshop is easier to control, but the quality is still raw. Generative Restyle (day to night, style changes). It looks impressive, but the intermediate frame costs 80 credits and loses quality.
Where Photoshop has no competition?
Text editing. Imagine a flat JPEG with a banner where one word needs to change. Photoshop converts the text in the image into a real text layer with a similar font. This is a beta and costs about 10 credits. GPT Image will also replace the text on request, but it redraws it as part of the picture. A layer can be edited as many times as you like, so Photoshop wins. Upscaling. Firefly Upscaler 2 scales an image up to 67 megapixels for 10–20 credits.
What looks odd and dated?
The author separately flagged what feels outdated: The built-in timelapse records only the canvas, and a regular screen recording captures more. Mockups: on a bottle, the design tears at the edges, and the old smart object with cylindrical warp handles it better. When asked to do frequency separation, the AI Assistant thinks for a long time and does nothing. Payment for every attempt: if you don't like the result, you pay again.
Ship AI Faster with Founders, Engineers & Weekly Architecture Reviews
Inside Vibe Club (hamidun.com/club) — weekly live masterminds with Zhemal Khamidun, Tech Radar, full LMS access, and 370+ production Claude Code skills.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.