Collection of GPT-image-generator 2.0 issues, bugs, and work-around tips (check first post)

[Tina_ChiKa]

It seems I didn’t make myself clear - I’m sorry, let me be more precise

For a start, I completely agree with you on this

@Rafaelis

The problem seems deeper than that.

This was just a example of my assumption as to where the deeper challenges might stem from:

@Tina_ChiKa

After all, there are many experiments with surrealism and fantasy where it isn’t always clear what “anatomically correct” might mean.

Surrealist works, for example, are highly rewarded.

My aim was not to play around with vague terms – but to use scientific terminology with a clear classification framework.

@Rafaelis

Better phrasing may improve some outputs, but I don’t think it solves the core issue.

I agree!

What you’re describing here

@Rafaelis

The model has always had a bit of a tendency to make arms too long, but in 2.0 it seems more noticeable, especially when limbs are partly hidden or pass behind something. It looks like a spatial coherence issue. The model loses track of the hidden part of the arm, then reconstructs it at the wrong length.

The same goes for hands. Specifying human hands or five fingers may help now and then, but it doesn’t reliably fix the underlying problem. Sometimes the instruction is ignored, and sometimes the model overcorrects somewhere else.

can also be observed in other fields.

A red kite has a distinctive fork in its tail feathers. That is why experts also refer to it as the ‘fork-tailed kite’. However, the models do not generate a red kite, but only a “bird of prey type”.

Unless the Latin term “milvus milvus” is used.

Here is the link to my post

Another example was that, a year ago, DALL-E could not generate roses, but by using an explicit botanical term for a specific rose, the image could be generated.

Here is the link to my post

@Daller You might still remember?

Since you’re specifically referring to ‘Anatomy’ here:

@Rafaelis

I’ve seen quite a few of these anatomy errors: feet with the wrong number of toes, thighs too long compared to the lower legs, heads either too large or too small for the torso, and so on. The model also seems very cautious with female busts, but that is not exactly new. It feels like another kind of overcorrection, where anything remotely prominent gets flattened or minimized.

I have written about a hypothesis, here is the link as well

@Rafaelis

Human anatomy depends on very precise spatial continuity, and that seems to be where a lot of these failures happen.

So I don’t think this is just a matter of users needing better wording. It may be that the training or tuning balance changed, and anatomy got worse as a side effect.

I agree with you 1000% on this, and I’ll take it a little step further .. by saying that this applies not only to human anatomy but also to structural differentiation.