AI and photo editing
Discussion
Today I tried getting Copilot to change the jumper on a person in a photo. No matter what instructions I gave it, the result was always about half the pixel dimensions. The jumper was fine, but the image was too small.
By chance I stumbled on Gemini. Aha, I thought, perhaps Gemini can do it. Changing a jumper can't be that hard for a technology that's taking over the world.
The conversation went like this:
By chance I stumbled on Gemini. Aha, I thought, perhaps Gemini can do it. Changing a jumper can't be that hard for a technology that's taking over the world.
The conversation went like this:
Simpo vs AI said:
Can you edit photos?
Yes, I can! Through the web interface, you can seamlessly modify existing images.
Here is what we can do together:
Targeted Local Edits: Add, remove, or change specific objects within a photo.
Background Replacements: Swap out the background of your image for an entirely new scene.
Image Blending: Combine multiple uploaded images into a single, cohesive photo while maintaining character consistency.
Guiding Edits: You can resize images or draw directly onto them to help guide my edits.
Text & Quality Fixes: I can improve text rendering and spelling inside images.
Feel free to upload a photo and tell me what changes you'd like to make!
Can you keep the resolution the same? eg if I give you a 3000 x 2000px image to edit, will the result also be 3000 x 2000 px? Copilot can't do this.
Unfortunately, no, I cannot keep the resolution exactly the same for an image that large.If you upload a 3000 x 2000px image, my image generation engine will downscale the final edited output.Here is how the limits break down depending on the tier you are using:Free Tier: The output is capped at 1K (1024 x 1024px).Gemini Advanced (AI Plus / Pro): The output is capped at 2K (2048 x 2048px). AI Ultra Tier: Supports up to 4K (4096 x 4096px). Why this happens: When you edit an image with AI, it doesn't just "paint" over the top like Photoshop. The system runs your original photo through a diffusion model to re-generate the modified sections, and that model has a strict maximum resolution pixel cap. If you have a 3000 x 2000px photo, an Advanced/Pro account will get you relatively close (compressing it down to fit within a 2048px limit), but it will still be a downscaled file. For professional, pixel-perfect resolution retention, traditional editing software or dedicated AI upscalers are still the best bet!
Is this simply to save AI processing power?
Actually, it s less about saving power on a whim and more about a hard technical bottleneck in how AI "thinks" about images.Traditional software like Photoshop sees an image as a static grid of individual pixels. It doesn't care if the grid is $1,000 \times 1,000$ or $10,000 \times 10,000$; it just changes the specific pixels you tell it to.AI doesn't work that way. To edit your photo, it has to convert the entire image into math (vectors) and pass it through a neural network. Here is why that limits the resolution:Exponential Memory Requirements: The VRAM (video memory) required by an image generation model doesn't scale linearly with resolution it scales exponentially. Doubling an image's dimensions quadruples the amount of memory needed to process it.The "Training" Limit: AI models are trained on specific dataset sizes (usually $512\text{px}$, $1024\text{px}$, or $2048\text{px}$). If you force a model to generate or edit at a resolution much higher than it was trained on, it gets confused and starts duplicating objects giving you a person with two heads or a landscape with two horizons.So, while it does save immense amounts of computing power, the restriction is primarily because current AI models physically cannot maintain coherence and quality if they try to process massive canvas sizes all at once.
So there you go - something you can do in seconds in Photoshop etc makes AI's head explode!Yes, I can! Through the web interface, you can seamlessly modify existing images.
Here is what we can do together:
Targeted Local Edits: Add, remove, or change specific objects within a photo.
Background Replacements: Swap out the background of your image for an entirely new scene.
Image Blending: Combine multiple uploaded images into a single, cohesive photo while maintaining character consistency.
Guiding Edits: You can resize images or draw directly onto them to help guide my edits.
Text & Quality Fixes: I can improve text rendering and spelling inside images.
Feel free to upload a photo and tell me what changes you'd like to make!
Can you keep the resolution the same? eg if I give you a 3000 x 2000px image to edit, will the result also be 3000 x 2000 px? Copilot can't do this.
Unfortunately, no, I cannot keep the resolution exactly the same for an image that large.If you upload a 3000 x 2000px image, my image generation engine will downscale the final edited output.Here is how the limits break down depending on the tier you are using:Free Tier: The output is capped at 1K (1024 x 1024px).Gemini Advanced (AI Plus / Pro): The output is capped at 2K (2048 x 2048px). AI Ultra Tier: Supports up to 4K (4096 x 4096px). Why this happens: When you edit an image with AI, it doesn't just "paint" over the top like Photoshop. The system runs your original photo through a diffusion model to re-generate the modified sections, and that model has a strict maximum resolution pixel cap. If you have a 3000 x 2000px photo, an Advanced/Pro account will get you relatively close (compressing it down to fit within a 2048px limit), but it will still be a downscaled file. For professional, pixel-perfect resolution retention, traditional editing software or dedicated AI upscalers are still the best bet!
Is this simply to save AI processing power?
Actually, it s less about saving power on a whim and more about a hard technical bottleneck in how AI "thinks" about images.Traditional software like Photoshop sees an image as a static grid of individual pixels. It doesn't care if the grid is $1,000 \times 1,000$ or $10,000 \times 10,000$; it just changes the specific pixels you tell it to.AI doesn't work that way. To edit your photo, it has to convert the entire image into math (vectors) and pass it through a neural network. Here is why that limits the resolution:Exponential Memory Requirements: The VRAM (video memory) required by an image generation model doesn't scale linearly with resolution it scales exponentially. Doubling an image's dimensions quadruples the amount of memory needed to process it.The "Training" Limit: AI models are trained on specific dataset sizes (usually $512\text{px}$, $1024\text{px}$, or $2048\text{px}$). If you force a model to generate or edit at a resolution much higher than it was trained on, it gets confused and starts duplicating objects giving you a person with two heads or a landscape with two horizons.So, while it does save immense amounts of computing power, the restriction is primarily because current AI models physically cannot maintain coherence and quality if they try to process massive canvas sizes all at once.
Gassing Station | Photography & Video | Top of Page | What's New | My Stuff