An image-removal tool's successful example is not a guarantee for your file. Results depend on the image, the treatment and the detector used to judge it. A small test can show that outcomes differ without explaining every cause.
In Deprint's research, one fixed processing chain received a negative Hive classification on a selected portrait. It remained positive on a desk scene and a bakery scene.
How far can one successful result travel?
A conceptual ladder of increasingly broad claims. Each step needs new evidence.
One file
What this means
Changing format or dimensions creates a new candidate, even when the filename is similar.
One scene
What this means
Twenty exports of one scene do not supply twenty independent examples of generalization.
One detector
What this means
Tuning on a detector and then scoring on it is a useful optimization test, but not an independent validation.
The same settings produced different outcomes
| Scene | Original control: Hive AI score | Processed output: Hive AI score | Output classification |
|---|---|---|---|
| Selected public portrait | 99.9% | 2.9% | Negative |
| Google-generated desk | 99.9% | 90.9% | Positive |
| Google-generated bakery | 99.9% | 99.9% | Positive |
The portrait control was resized and JPEG-encoded to match its final delivery format. Desk and bakery controls were positive in the recorded checks. The chain combined redraw, a JPEG step, detail restoration, camera-style processing and export. Its settings were held fixed for the scene comparison.
This is three reused study scenes, including an already-promising portrait. It is not a representative one-in-three success rate. The scenes differ in content, dimensions and source history, so the comparison does not isolate which difference caused the failures.
Was the successful image simply smaller?
The portrait's final export contained about 1.38 million pixels; the desk and bakery exports had about 4.3 million each. A follow-up resized both failed examples to the portrait's pixel count, preserving each aspect ratio. Their original controls received the same resize and JPEG export.
| Size-matched scene | Detector used | Original score | Processed score |
|---|---|---|---|
| Desk | Sightengine GenAI | 99% | 99% |
| Bakery | Hive AI | 99.9% | 99.9% |
Reducing the delivered size did not resolve these failures. The two rows use different detectors and must not be averaged. This test also does not establish that all possible resizing methods have the same effect.
More processing did not rescue the difficult scene
Stronger redraws were tried on the desk. They changed more of the image but remained Hive-positive. That matters when you are tempted to keep increasing a “strength” slider: visible damage can grow without producing the result you want.
Read the stronger-processing comparison before assuming that more treatment is better.
How to judge a tool's examples
Look for results on images similar to your own, but do not stop there. A useful test also shows untreated originals, failed outputs and the exact detector used. Be cautious about results collected only from examples chosen after the scores were known.
A portrait, a product label and a wide interior scene place different demands on preservation. Even if all three pass a detector, their usefulness still depends on whether important details survive.
If your image fails, keep the failed output and original. Check that you are using the same detector and the actual exported file. Do not assume a tool's success message means it checked the result externally.
A successful demo and a dependable tool answer different questions
A demo asks whether a favorable output can be produced for an example. A reliability study asks what happens across new inputs under a defined procedure. Both can be useful, but they support different promises.
If a portrait was chosen because it already looked promising, it is especially important to test fresh scenes afterward. Repeating that portrait with different seeds, file types or export sizes can teach you about the portrait. It does not give the same evidence as new independently selected scenes.
| Collection | What it can help answer | What it cannot establish alone |
|---|---|---|
| Many variants of one portrait | Sensitivity to settings on that portrait | Reliability across portraits |
| Several scenes from one generator | Variation within that selected source | Performance on every generator |
| New scenes held aside before tuning | Some evidence of transfer | Universal performance outside the sample |
| Only successful outputs | What a favorable case looks like | How often failure occurs |
Why the scene differences are clues, not explanations
The portrait, desk and bakery differ in content, framing, dimensions and source history. It is tempting to conclude that faces are easier or that interiors are harder. This comparison does not isolate those categories, so it cannot establish that rule.
A causal explanation would need a more controlled design. For example, you could keep the method fixed and independently vary a defined property across a planned sample. You would also need enough independent images to distinguish a consistent pattern from a few unusual cases.
Until then, the honest explanation is that the method did not generalize to these two scenes. That is already useful information for a reader deciding how much confidence to place in one attractive demonstration.
What to do when your own file fails
Confirm the target first. A watermark-negative result and a visual-classifier failure can coexist. Next confirm that you tested the actual downloaded output, not the original, an intermediate or an editor thumbnail. Check the tool and observation date.
Then inspect quality. If the first attempt changed text or faces, increasing processing strength may make the image less useful. Keep the failed output and return to the untouched source for a different comparison, rather than building an undocumented chain of edits on top of it.
For a recurring workflow, collect a small set of your real input types before committing to a method. Include difficult examples and keep failures. The goal is to learn whether the method meets your needs, not to reproduce a marketing example at any cost.
Our current tests suggest which follow-ups deserve attention. They do not identify a guaranteed best setting for a new image. The test-design guide shows how to turn that uncertainty into a more useful evaluation.
What these results do and do not establish
The October 8–9, 2026 observations demonstrate variation across these scenes. They do not establish which content category is easiest, which generator is hardest, or what fraction of future images will pass.
A stronger follow-up would freeze the treatment before testing new images, cover several types of content, and publish the full set of outcomes. It would also measure image damage alongside detector results. Read the frozen study methods and data.
For the related distinction between watermark and visual checks, read does removing SynthID bypass AI image detectors?.