GPT Image 2 produces plastic-looking faces in product edit

ChatGPT Bugs
, , ,
1

Hello team,

I am reporting a serious quality problem with GPT Image 2 when editing product images.

I use image editing for a professional catalog. The workflow requires preserving the exact garment, pose, materials, colors and product construction while changing only a specific garment color or pose.

With GPT Image 2, the edited images frequently produce faces that look visibly plastic, overly smooth and artificial. The face loses natural skin texture and realistic facial detail, even when the prompt explicitly requires preserving the original model and changing only the garment color.

This does not happen to the same extent with ChatGPT Image Latest. Image Latest generally preserves a more natural-looking face and better visual continuity with the reference image.

The problem is especially serious for commercial product photography because the face is part of the perceived quality and identity of the product image. Strong product-lock instructions and repeated constraints do not reliably solve it.

Attached is a contact sheet showing five face crops from GPT Image 2 product-edit outputs. The complete product images were intentionally cropped so we can evaluate the facial quality issue directly.

Could the team please clarify:

  1. Is this a known regression or limitation in GPT Image 2 editing?

  2. Is there a recommended setting, mask strategy or workflow that reliably preserves realistic facial identity and skin texture?

  3. Will Image Latest remain available for workflows where facial fidelity is more important than the newer model’s other capabilities?

  4. If Image Latest is planned for deprecation, could this face-quality gap be addressed before it is removed?

For professional catalog, fashion and e-commerce workflows, realistic face preservation is essential. At the moment, GPT Image 2 is not consistently reliable enough for these production use cases.

2

Hi @vikgravina,

Firstly I am just another user here.

Secondly, I understand why this is a problem for legitimate catalog workflows. At the same time, seeing children’s faces used in the example immediately made me think about the other side of realistic facial-identity preservation: especially where minors are concerned, it also raises serious deepfake and safeguarding concerns.

I would rather see that they improve the image-editing quality, but keep restrictions that prevent people from using it to create convincing deepfakes, especially of minors.

3

Hi Larisa, thank you for raising this.

The exact same issue occurs with adults as well.

4

Do you have some examples and prompts you can share? Otherwise, it’s a bit challenging to help you get better results.

5

Yes, absolutely. I’m attaching a representative image showing the issue on adult faces.

I have been working with GPT Image 2 since it was introduced in April. From the very beginning, in the type of catalog-image work I do, this problem has been consistently present during image editing.

I have tested thousands of generations with many different prompts. Sometimes the effect is weaker and sometimes stronger, but the result is essentially always the same: an unnatural, plastic-like layer appears over the face and other exposed skin areas.

It is most obvious on the face because the skin is fully visible. On clothing it can be less noticeable because the garment texture hides part of the problem. However, the artificial facial rendering is very clear.

Compared with ChatGPT Image Latest, Image 2 does not produce the same level of natural facial realism. Image Latest may interpret some edits differently, but its results look significantly more realistic and photographic.

Attached representative example:

The issue persists even when the prompt explicitly requires preserving the exact face, identity, skin texture, and all details except the garment color.

gpt-image-2-issue
gpt-image-2-issue1920×1920 217 KB

PROMPT 1 — Minimal

Change the garment color from black to deep burgundy red. Create a realistic studio photograph.

PROMPT 2 — Basic Preservation

Change the garment color from black to deep burgundy red. Keep the person, pose, hairstyle, garment design, lighting, framing, and background consistent with the reference image. Make no other intentional changes.

PROMPT 3 — Professional Image Edit

Edit the attached photograph by changing only the garment color to deep burgundy red. Maintain the original composition, subject, pose, proportions, garment shape, material appearance, lighting, shadows, accessories, shoes, and background.

PROMPT 4 — Controlled Color Replacement

Replace the black color of the garment with deep burgundy red while keeping the original photograph visually consistent. Preserve the subject’s appearance, pose, hairstyle, facial expression, garment construction, seams, texture, lighting, framing, accessories, shoes, and background. Do not add or remove anything.

PROMPT 5 — Structured Production Prompt

Task: recolor the garment from black to deep burgundy red.

Input: the attached photograph.

Expected result: a realistic studio photograph with the garment shown in deep burgundy red.

Keep consistent: subject, facial appearance, hair, body, pose, proportions, garment design, fit, materials, seams, texture, lighting, shadows, framing, accessories, shoes, and background.

Scope: change the garment color only. No additional modifications.

6

Thanks for sharing the example and prompts, and sorry the edits aren’t preserving the look you need. We’ll pass this along for review. We don’t have a confirmed fix or timeline to share for this specific issue.