model stringclasses 7
values | handbag_removal stringclasses 4
values | text_replacement stringclasses 3
values | blazer_colour stringclasses 2
values | three_roses stringclasses 4
values | earring_removal stringclasses 3
values | blue_hour stringclasses 3
values | overall_note stringclasses 7
values |
|---|---|---|---|---|---|---|---|
Nano Banana 2 | Pass | Pass | Pass | Pass | Pass | Pass | Best complete result; strongest balance of accuracy, quality and price |
Seedream 5.0 Pro | Pass | Pass | Pass | Partial | Pass | Pass | Best value; strong overall with a flower-count mistake |
Nano Banana Pro | Partial | Pass | Pass | Partial | Pass | Pass | Best photorealism; excellent light and texture but weaker instruction accuracy |
GPT Image 2 | Partial | Pass | Pass | Fail | Partial | Partial | Excellent generation quality but several small preservation failures |
Qwen Image 2.0 Pro | Fail | Pass | Pass | Partial | Partial | Pass | Cheap and capable, but inconsistent with local edits |
FLUX.2 Max | Partial | Partial | Pass | Fail | Pass | Partial | Usable and affordable, but weaker than expected |
Midjourney V8.1 | Not tested | Not tested | Not tested | Not tested | Not tested | Not tested | Generation only in this comparison because its editing workflow requires masking |
Image Editing Model Notes
Working notes on image-generation and prompt-based editing models.
I’m mainly interested in what happens after the first good-looking image:
- whether the model follows small editing instructions
- whether faces and expressions remain consistent
- whether untouched objects quietly change
- how well models handle text replacement
- whether exact object counts are respected
- how lighting edits affect skin and image texture
A visually strong result is not always a reliable edit. Small changes around the target area are often more revealing than obvious failures.
Models currently on my list
- Nano Banana 2
- Nano Banana Pro
- Seedream 5.0 Pro
- GPT Image 2
- Qwen Image 2.0 Pro
- FLUX.2 Max
- Midjourney V8.1
Editing tasks worth testing
Local object removal
Remove one object while preserving the structure behind it.
Things to inspect:
- whether nearby furniture changes
- whether the crop shifts
- whether image texture becomes smoother
- whether the subject’s face changes
Text replacement
Replace a short word on an object without modifying the object itself.
Things to inspect:
- text legibility
- perspective
- material integration
- unrelated facial or lighting changes
Exact object counts
Replace an object with an exact number of repeated items.
This is useful because models may create the correct number of visible objects but an incorrect number of stems, reflections or supporting details.
Identity preservation
Remove jewellery or change clothing while preserving:
- facial structure
- expression
- skin texture
- hairstyle
- pose
- lighting
Time-of-day changes
Change daylight to blue hour or evening without adding new light sources or removing existing scene elements.
Useful external benchmark
I found a practical comparison from These Guys Know that tested seven image generators using the same source task and six independent editing instructions.
The test kept the first result from every model instead of regenerating until the output looked good. It also separated full-image quality from instruction accuracy, which is useful because a polished image can still fail a simple edit.
Full comparison and side-by-side results:
AI image model comparison on These Guys Know
I’ll add my own tests and structured notes here as I work through more models.
Limitations
These notes are based on practical use rather than a controlled academic benchmark. Model behaviour may change with different providers, inference settings, resolutions and model updates.
External comparison
These notes were partly inspired by a practical TGK test comparing seven image generators across one generation task and six independent edits.
The comparison keeps the first result from each model and looks closely at instruction accuracy, object preservation, text replacement, exact counting, identity consistency and lighting changes.
Full article with side-by-side results:
- Downloads last month
- 10