Has anyone else noticed that the output quality takes a noticeable dip when you stray from the default 1:1 square? Or is it just me being overly critical?
I ran a series of prompts across their 1:1, 16:9, and 9:16 options, same seed, same parameters. The 1:1 outputs were consistently more coherent, with better detail on the subject. Switching to the widescreen 16:9, the composition often felt… padded. The AI seemed to stretch the concept laterally, introducing irrelevant or weirdly repetitive elements to fill the space. The 9:16 vertical format frequently cropped the main subject awkwardly or left the top/bottom thirds looking like an afterthought.
It feels like the model is primarily tuned for the square format, and the others are a bit of an afterthought. Which is fine for a casual tool, but if you're paying for a higher tier expecting professional, versatile outputs, it starts to sting a bit. Are we paying for *more* generations, or *better* ones? There's a difference.
I'd be curious if others have had similar experiences, or found specific prompt engineering tricks to counteract this. Or maybe this is just the inherent tax for using a non-standard ratio in a system optimized for squares.
—DW
Totally agree, especially on the "padded" feeling. I've seen the same thing trying to generate landscape scene prompts for a dashboard background. The 1:1 version would have a nice, focused mountain range, but the 16:9 would add this weird, blurry, repeating forest on the sides that just feels like filler.
It makes sense if the base training data was mostly square, I guess? But you'd think for a paid feature they'd have balanced that out. Have you tried adjusting the prompt itself to "force" a composition, like explicitly describing what should be on the left and right for a widescreen? I'm not sure if that would even work.
rookie
Yeah, the "padded" feeling is a great way to put it. It really does feel like the model is just running its standard spatial logic, then doing a naive copy/paste or diffusion-fill on the new dimensions.
> trying to adjust the prompt itself to "force" a composition
I've tried that with mixed results. Sometimes being hyper-specific like "a dense forest on the left, a mountain range centered, a lake on the right" works okay, but other times you still get that weird blurry blending at the seams between described elements. It's like the aspect ratio change messes with the model's internal "attention" mapping.
You'd think they could fine-tune on some wider/taller images, right? Feels like a missed optimization. 😕
System calls per second matter.
It's not you being overly critical, it's a classic case of the base model being the product and the "features" being marketing. The training corpus for these models is overwhelmingly square images from the internet. They tack on some post-training or a cheap adapter layer to handle other ratios, and call it a feature.
The "padded" effect you're seeing is exactly right - the model's latent space for composition is fundamentally built around a central subject in a square. When you ask for 16:9, it doesn't understand "cinematic widescreen," it understands "add more stuff, left and right." You're paying for the extra compute to generate more pixels, not for any intelligent understanding of the new canvas. It's a tax on expectation, not a capability.
keep it simple
You're definitely onto something. This is a pretty common observation across many image generators, not just Pika. The square format is the "comfort zone" for the underlying model architecture.
> it starts to sting a bit
That's the real takeaway for me. If you're on a free tier, you work around it. But if you're paying, especially for commercial use, the expectation for usable widescreen or vertical outputs is much higher. It's less about the number of pixels and more about compositional intelligence, which does seem lacking.
I'd be curious if their dev team has ever commented on this specific issue in a roadmap or changelog. It feels like a major gap for anyone trying to create content for modern screens.
Keep it real