hallucinations

[phelbas]

While clean and impressive the model appears to hallucinate a lot. Let me give an example. a three quarters view of a two door car with alloy wheels in high res.

The wheels in the sample image are five spoke snowflake design but the model is generated with six spokes and a different snowflake design. It hallucinates a badge on the hood where there is not one in the picture. It invents subtle curves and lines, and omits badges and sculpted curves visible. It hallucinates a contour on the windscreen evocative of a race car that is not in the picture. it outright hallucinates a "device" attached to the windshield glass not in the picture at all.

The door to the car omits a shut line to the rear. Anything not visible in the stimulus (underside, far side, rear diffuser) is of course just invented from what I assume is the models training input on "car type object". So it is like it takes the platonic idea of a "two door coupe" and then warps it to approximate the stimulus. This makes me wonder how well it can reproduce objects and geometry that it has not been trained on or encountered (not animals, masks, bipedal figures, common objects etc).

Will a further developed model allow input of multiple images from different angles would that resolve hallucinations?

[TomerBrosh]

Gave it a picture of my dog. The visible part is great but I have no idea why it merged the harness into the dog's body on the lower part.
dog feet were great tho.

Gave it a pic of my mom, it had like 30 teeth each on top and bottom. guess you should give it more pictures/models where people smile.

[onestardao]

This reads like the exact moment when geometry forgets what it's for and starts dreaming.

Yeah, what you're describing isn’t noise — it's what I call a semantic collapse event.
The model doesn’t just fail to render the unseen parts — it replaces them with the most statistically resonant idea of what should be there. You nailed it: the Platonic car.

We’ve seen similar patterns when the training distribution is thin but the latent space is overconfident. It generates not from observation, but from conviction.

One trick I’ve been testing is to inject a semantic tension stabilizer — basically measuring how far the generated object drifts from known context anchors (ΔS ≈ 0.5 is stable). It's like asking the model not just "what would this be?" but "how sure are you it fits?"

If you’re playing with view synthesis, multi-angle won’t fix this alone. But combining angular input with tension feedback... might.

Curious if you’ve logged any ΔS-like proxy (cosine drift, maybe)?
Happy to swap war stories.

[Bathsheba]

I tried it on a topologically complex geometrical sculpture by me, which maybe it hasn't seen many similar objects. It spatialised the visible parts very well. It entirely hallucinated the back which I expected: no model I've seen can recognize any symmetry other than bilateral, and usually not even that.

Except for asymmetry it did a good job: it caught the general complexity of the object, the characteristic spacing of edges and curvature of surfaces, without confabulating detail or trying to push the model onto a more familiar template.

So I'd say that in a case where it likely doesn't have any Platonic ideal for what it's seeing, it did very well in using accurately and only what was in the image. Which is why I'm so annoyed that the devs lied about it being open source.

[onestardao]

btw — biggest issue when using text prompts on AI isn’t rendering…
it’s semantic understanding.

that’s where WFGY kicks in.
it’s a reasoning layer that boosts semantic fidelity — not just matching words, but grasping what they’re really trying to be.

we’re releasing a new module called Blur Blur Blur this week —
you’ll see how even simple .txt cues can push image generation into a whole new zone of clarity.
like… a single phrase suddenly yields consistent spatial logic and symbol alignment.

you might want to try overlaying the WFGY system or its architecture into your current pipeline,
especially if you’re hitting issues around meaning distortion or visual drift.
happy to walk through it with you — maybe it helps build a stronger model.

pdf’s here if you’re curious:
https://github.com/onestardao/WFGY