The npot property shouldn't meaningfully change the quality. In fact, I'd even consider removing it since WebGL 1 support has been dropped. So quite surprised to see it used this way.

I haven't tested it myself, so I do take your word that there's a perceivable difference. My guess is that it's caused by the minFilter being set to gl.LINEAR instead of gl.LINEAR_MIPMAP_LINEAR (material.js#L36). This introduces more aliasing, especially at larger distances or at off angles. In this viewing configuration that probably results in a (slightly) sharpened image which helps with the outlines.

Though I wouldn't recommend using it this way. At the very least it would be cleaner to directly set the minFilter instead of using npot: true for the same effect. A better approach would be to set the texture lod bias, but Three.js sadly doesn't exposes this property.

Yes that's exactly the lines

texture.wrapS = THREE.ClampToEdgeWrapping;
texture.wrapT = THREE.ClampToEdgeWrapping;
texture.magFilter = THREE.LinearFilter;
texture.minFilter = THREE.LinearFilter;

that improves the sharpness. I have those lines on a custom component that creates a sphere geometry and material for 360 image.

Should we set minFilter instead?

Except for repeat, offset, anisotropy, setting other texture properties is currently not exposed to the material component schema.
I don't know what that npot: true was doing in the past with WebGL1, today on WebGL2 it's just setting those 4 texture properties.
That may not be a good name now to mean just setting linear filter, but that's a property that is used in the wild, so I guess we'll keep it as backward compatibility.

WebGL 1 placed various restrictions on non-power-of-two (npot) textures. Including the wrapping and filtering modes that can be used with such textures. With WebGL 2 there are no restrictions for npot textures.

I'd say we should start exposing the filtering methods on the material component. Indirectly setting it through the npot property is unintuitive. Not to mention that it's unclear if npot is used for legacy reasons (WebGL 1 support), to enable clamping or to use linear filtering/avoid mipmaps.

I’m good with exposing the new properties and deprecating npot

Okay. Feel free to still merge this and we can work on that later. I don't plan to work on it right now.

Thanks