Can DALL-E Restore Old Photos?

DALL-E 3, OpenAI's image generation model, is capable of producing remarkably detailed and realistic images from text descriptions. If you're wondering whether it can restore old, faded, or damaged family photographs, here's the direct answer:

DALL-E cannot restore old photographs. It generates new images — it doesn't repair or reconstruct information in existing ones.

---

What DALL-E Does

DALL-E 3 is a text-to-image model integrated into ChatGPT. You describe an image in text, and DALL-E generates it. OpenAI's safety filters also mean it won't generate realistic portraits of real people based on descriptions.

Features that seem adjacent to restoration:

Image generation from description: "A 1940s portrait of an elderly woman in formal dress" — DALL-E can generate this, but it's a fictional person, not the individual in your family photo.

Inpainting (via API): DALL-E's API supports image editing — providing an image plus a mask, and DALL-E fills the masked region with generated content. For old photos, this replaces damaged areas with AI-generated content that plausibly fits the composition — not the original content that was there.

Variations (DALL-E 2): Creates variations of an input image. Generates images inspired by the original — not repairs to it.

---

Why DALL-E Can't Restore Photos

It generates, it doesn't reconstruct. Every output from DALL-E is a new image created from learned statistical patterns. When you provide an old photo as reference, DALL-E uses it as inspiration to generate something new — the output is generated, not restored.

The face problem: For a faded 1965 portrait, the specific person's facial features (in degraded but recoverable form) exist in the original scan. CodeFormer was trained on historical face degradation patterns to reconstruct that information. DALL-E was trained to generate plausible faces — it would produce a fictional face that looks compositionally right, not the original person's face recovered.

OpenAI's safety restrictions: DALL-E explicitly limits generation of realistic portraits of real individuals. Even if DALL-E could technically infer the original face, safety systems would restrict this use case.

---

GPT-4o Image Editing (April 2025 Update)

OpenAI's GPT-4o gained image editing capabilities in 2025. Users can upload images and ask GPT-4o to make edits. Testing on old photos shows:

• GPT-4o can brighten images, adjust contrast, and apply simple corrections
• For scratches, it attempts to fill them — generating content to cover the damage, not recovering original information
• Face reconstruction follows the same generative pattern: new faces, not original faces recovered

GPT-4o's image editing is more capable than DALL-E 2's inpainting, but it's still generative AI applied to editing — not purpose-built restoration.

---

Comparison

| Task | DALL-E / GPT-4o | ArtImageHub | |------|-----------------|-------------| | Generate a new portrait from description | ✅ | — | | Restore original face detail from old photo | ❌ Generates new face | ✅ CodeFormer | | Fix systematic fading/yellowing | ❌ No | ✅ GFPGAN | | Remove scratches (recover original) | ❌ Generates fill content | ✅ AI pattern recognition | | Colorize B&W accurately | ❌ Generates plausible colors | ✅ Yes | | 30-second results | ✅ | ✅ |

---

When DALL-E Is Useful Near Old Photos

Creative reimagining: If you want an artistic interpretation of an old family portrait — in a specific style or era — DALL-E can produce interesting creative results. With the understanding that it's creative work, not restoration.

Background generation: If an old portrait has a completely destroyed background, DALL-E's inpainting (via API) could generate a period-appropriate replacement — a creative fill, not a factual one.

These are creative applications, not restoration.

---

What Actually Restores Old Photos

CodeFormer — trained on historical photographic degradation patterns to reconstruct original face detail. This is what runs in ArtImageHub.

GFPGAN — systematic correction of fading, yellowing, and color shift specific to aging photographic paper. Not general contrast adjustment.

Real-ESRGAN — AI upscaling designed for real-world degraded images, not generative upscaling.

The difference between these restoration models and DALL-E: restoration models reconstruct what was originally there; DALL-E generates what plausibly could have been there. For family history, that distinction matters.

---

Restore your old family photos at ArtImageHub — $4.99 one-time →

Results in 30–90 seconds · HD download · 30-day guarantee

---

Related

• Can ChatGPT Restore Old Photos? — full ChatGPT/GPT-4o analysis
• Can Midjourney Restore Old Photos? — Midjourney analysis
• Can Stable Diffusion Restore Old Photos? — open-source AI analysis
• Best AI Tools for Old Photo Restoration in 2026 — 7-tool ranked comparison