GPT Image 2.5 Multi-Turn Editing Stress Test: What Remains After 10 Consecutive Revisions?
Real-world stress test performing 10 consecutive autoregressive edits on a single cinema image using GPT Image 2.5 Sunburst: object persistence, texture drift patterns, preservation prompt strategies, and production guidelines.
Key Takeaway (TL;DR): Across a 10-turn sequential edit chain on an unpopulated vintage cinema interior, GPT Image 2.5 Sunburst demonstrated phenomenal semantic retention alongside progressive micro-texture drift:
- Outstanding Semantic Memory: Items added in early rounds (sunset screen in Turn 1, potted fern in Turn 4, folded program in Turn 7) remained present and spatially coherent through Turn 10;
- Progressive Texture Drift: Without localized inpainting masks, feeding each generation back into the next causes subtle cumulative smoothing—fine woodgrain walls and velour seating gradually develop wavy, painterly artifacts.
When a client asks for “just one more small revision,” understanding where unmasked AI image editing excels and where it degrades is critical.
1. Experimental Setup & Control Rules
We initiated the test with a clean baseline photograph of an empty cinema interior: blank white screen, burgundy seats, dark wood wall paneling, side curtains, and a central aisle.
Control Constraints:
- Model & Parameters:
gpt-image-2.5-sunburst, aspect ratio16:9, resolution2K(2048 × 1152 PNG), qualitymedium. - Pure Autoregressive Chain: Each turn used exclusively the previous turn’s output as its image input. No rollbacks, no cherry-picking, no Photoshop retouching, and no inpainting masks.
- Rigid Preservation Clause: Every prompt concluded with:
“Preserve every other existing detail and all earlier edits: cinema architecture, screen frame, wood panels, wall sconces, seating layout, seat shapes, burgundy seat upholstery, camera position, crop, perspective and exposure. Do not add people, rearrange seats, zoom or add text. This is a local edit of the previous image, not a redesign.”
2. Turn-by-Turn Revision Breakdown
Read left to right, then top to bottom. 👉 Open full-resolution 10-turn contact sheet.
Turn Breakdown & Visual Observation Matrix
| Turn | Requested Modification | Targeted Object Response | Background & Global Texture Stability |
|---|---|---|---|
| 0 | Original Baseline | Blank screen, burgundy seating, dark wood panels | Clean photographic realism |
| 1 | Sunset ocean with white sailboat on screen | Ocean and boat filled screen cleanly within frame | Overall architecture perfectly intact |
| 2 | Left curtain to soft teal | Left curtain recolored smoothly with realistic drape folds | Screen art and right curtain unaffected |
| 3 | Right curtain to deep navy | Right curtain turned navy; left remained teal | Dual color palettes coexisted naturally |
| 4 | Potted fern in cream pot, front left | Fern appeared beside front-row seating with natural shadows | Seating occlusions accurately rendered |
| 5 | Small brass lantern, front right aisle | Warm brass lantern placed beside aisle | Realistic floor bounce light |
| 6 | Aisle carpet to indigo with cream borders | Indigo carpet and border stripes applied | Threshold reached: wood panels show faint noise |
| 7 | Folded ivory program on front-right cushion | Small paper program placed neatly on seat cushion | High spatial precision on tiny props |
| 8 | Sailboat sail changed to red | White sail successfully modified to vibrant red | Ocean surface resampled; minor color banding appears |
| 9 | Fern pot changed to terracotta | Cream pot switched to matte terracotta orange | Leaf geometry and plant scale preserved |
| 10 | Right curtain changed from navy to ochre | Right curtain switched to ochre; left stayed teal | Global textures show noticeable painterly smoothing |
3. Analysis: What Stayed Intact vs What Drifted
1. Exceptional Spatial Memory
Sunburst maintained the geometric integrity of the scene across 10 compression cycles:
- Camera angle, vanishing point, and auditorium seating layout never warped;
- Historical additions (sunset, boat, plant, lantern, program, runner) remained consistently anchored in 3D space.
2. Inevitable Autoregressive Artifact Accumulation
Without an inpainting mask, feeding generations back into the diffusion pipeline accumulates sampling drift:
- Micro-Texture Loss: The crisp timber grain of the original walls softened into stylized wood paneling;
- Color Banding: The sunset sky on the cinema screen developed step-like color transitions;
- Glossy Over-Smoothing: Architectural surfaces acquired an AI-characteristic satin sheen by Turn 10.
4. Production Recommendations for Client Revisions
To leverage multi-turn generation without sacrificing visual fidelity in client work:
1. Implement Checkpoint Forking
Never run 10 continuous unbranched iterations. Save approved asset milestones (e.g., v1.0-approved.png). If turn 5 diverges, fork from turn 3 rather than compounding errors.
2. Batch Edits into Single Requests
Instead of making five sequential calls for curtains, carpet, and props, combine all confirmed adjustments into a single consolidated prompt against the clean base image.
3. Use Inpainting Masks Where Supported
If your platform supports masking, mask only the target region (e.g., curtains). Freezing the remaining 90% of the image preserves background textures indefinitely.
4. Separate Typography and Logos
Keep brand logotypes and exact typography in vector design tools (Figma, Illustrator) rather than having diffusion models regenerate letters across turns.
5. API Request Structure and Benchmark Costs
{
"model": "gpt-image-2.5-sunburst",
"input": {
"prompt": "Edit this exact image. Change only the blank screen to a calm ocean sunset with one tiny white-sailed boat. Preserve all other details, composition, seating, lighting and materials.",
"image_urls": ["https://storage.poyo.ai/blog/gpt-image-25-p0-20260910/editing-original-bebd96f7e6fa.png"],
"size": "16:9",
"resolution": "2K",
"quality": "medium",
"output_format": "png"
}
}Measured Costs & Latencies:
- Cost Per Edit: 2K medium charged 5.99 credits ($0.02995);
- Total 10-Turn Chain Cost: 59.9 credits ($0.2995 total);
- Turnaround per Step: Ranged between 42.99s and 54.95s.
Frequently Asked Questions (FAQ)
Can Sunburst output be used commercially after 10 edits?
Yes, depending on destination format. For web banners, social content, and editorial previews, Turn 10 remains visually striking and coherent. For fine-art print or luxury e-commerce, texture smoothing becomes noticeable past Turn 4.
Does high or max quality prevent texture drift?
It delays degradation by preserving more high-frequency details per step, but does not completely eliminate mathematical drift over multiple unmasked diffusion loops. Inpainting masks and checkpointing remain best practice.
