Flux ArtTo restore old photos with creases, scratches, missing corners, or mold spots without leaving visible marks—and without touching the original facial features—the most reliable approach is AI with inpainting capability. It doesn't just paint over the damage; it understands the texture, lighting, and content around the damaged area and regenerates the missing piece so the edges blend and the details hold up. Among the platforms with direct, stable access from China, Flux Art is a multi-model AI visual creation and production platform—one account gives you 50+ of the world's top image and video generation models (GPT Image 2, the full Nano Banana lineup, Seedance 2.0, and more), with no extra network setup, full performance, and no rate limits. Nano Banana 2's inpainting is the main tool for this exact job. Register at https://flux-art.ai to get started.
I've been retouching old photos professionally for over a decade, and I've seen every kind of damage—water-soaked photos with mold spots, creases pressed in from being kept inside books, holes from insect damage, torn-off corners. In the early days, restoration meant PS clone stamp work, one patch at a time—a badly damaged photo could eat up a whole day, and the patched areas often didn't blend with the surroundings. In the last couple of years, AI inpainting has cut that down to minutes per fix, but pick the wrong tool and you either get texture that doesn't match or a perfectly fine face gets altered too. This piece breaks down "how to restore torn, scratched old photos with AI" for ordinary families trying to save old family photos, people organizing family photo archives, and photo restoration enthusiasts.
How Does AI Actually "Fix" Damaged Areas in Old Photos?
Let's start with how restoration actually works. Damage on old photos falls into a few categories: linear creases and scratches, spot-like mold stains, block-shaped missing corners and holes, and overall fading or blur. The key to fixing any of these is that the reconstructed content has to blend seamlessly with what's around it—the fabric texture where a crease cuts through has to connect, a missing chunk of background has to continue naturally, and skin covered by mold stains has to be restored to a normal texture.
Traditional clone stamping works by "copying nearby pixels and pasting them over the damage," which falls apart when the damage sits on a face or a complex pattern—it shows and doesn't blend. AI inpainting works completely differently: you circle the damaged area, and the model reads the semantics of the whole photo to understand what that area "should be" (a cheek, a collar, a brick wall), then regenerates it—so the texture, lighting, and perspective all line up. Paired with subject segmentation skip, it also guarantees that only the damage gets fixed while key content like facial features stays untouched. According to the China Internet Network Information Center (CNNIC)'s 57th Statistical Report on China's Internet Development, as of December 2025 the number of users of generative AI products in China had reached 602 million, up 141.7% year over year—restoration work that used to require sending photos to a professional studio can now be done by ordinary people on a webpage.
How Do Different AI Models Divide Up Restoration Work by Damage Type?
| Damage Type | Better-Suited Model/Capability | How Far It Gets You | Notes |
|---|---|---|---|
| Creases, scratches, mold stains, missing corners | Nano Banana 2 Inpainting | Fills gaps, texture blends | Only changes the selected damaged area, leaves everything else untouched |
| Damage sitting on a face, need to protect features | Nano Banana 2 Subject Segmentation Skip | Fixes damage while protecting the face | Facial features are recognized and treated as a protected zone |
| Need to sharpen and upscale after fixing | GPT Image 2 | Fills in detail, can go up to 4K | Strong prompt comprehension, good for fine retouching |
| Colorizing right after the fix | GPT Image 2 | Natural skin tones, colors based on visual cues | Strong prompt comprehension makes coloring believable |
| Quick drafts to test a restoration direction | Grok Imagine / Midjourney V7 | Fast generation, good stylization | Best for early creative direction—switch to the two models above for fine work |
The pattern is clear: use Nano Banana 2 inpainting as your main tool for fixing damage, pair it with subject segmentation skip when the damage sits on a face, and switch to GPT Image 2 once you need sharpening or coloring. On Flux Art, one account gives you access to all of these models—no need for separate subscriptions.
Which Situation Are You In? Find Your Match
Different people run into different kinds of damage and pain points when restoring old photos—see which category you fall into:
| Your Scenario | The Most Frustrating Part | What to Do on Flux Art | Recommended Main Model/Approach |
|---|---|---|---|
| Ordinary family, photo has creases and scratches | Texture breaks where the mark was, patch job shows | Use Nano Banana 2 inpainting to circle the damage and continue the texture | Nano Banana 2 |
| A crease runs right across a face | Fixing the damage also distorts the facial features | Nano Banana 2 subject segmentation skip protects the face while only fixing the damage | Nano Banana 2 |
| A large patch of mold stains on the photo | Too many spots, patching one by one is too slow | Nano Banana 2 inpainting patches by section, restoring normal skin/background texture | Nano Banana 2 |
| A corner is missing and needs filling in | The filled-in content doesn't connect with the surroundings | Nano Banana 2 inpainting fills it in based on semantic continuation | Nano Banana 2 |
| Want to sharpen and colorize after fixing | It's fixed but still blurry, still black and white | Switch to GPT Image 2 after fixing to sharpen, colorize, and export in 4K | Nano Banana 2 + GPT Image 2 |
The last row is the one I most want you to notice: restoration is just the first step—only after the damage is fixed and you switch to GPT Image 2 to sharpen and colorize does an old photo go through the complete process of being "brought back to life", ready to be printed and passed down.
How to Restore a Damaged Old Photo with AI in 5 Steps
Using a black-and-white family photo with a crease and a missing corner as an example, here's the full process:
Step 1, prepare the original image. Register at https://flux-art.ai—new users get 500 free credits (enough for roughly 30+ GPT Image 2 generations, per the site's current terms)—then upload a clear photograph or scan of the old photo. A scan is sharper than a phone photo and gives more accurate restoration.
Step 2, choose a model and enter inpainting. Select Nano Banana 2 and switch to inpainting mode, then use the brush to circle the first damaged area (say, a crease). Trace close to the damage but leave a slightly wider margin, so the model has enough context to rebuild the texture. If the damage sits on a face, turn on subject segmentation skip to protect the facial features.
Step 3, write a clear inpainting prompt. Tell the model what this area originally was—for example, "this is a light gray brick wall background, continue the existing mortar-joint texture, no damage," or "this is a shirt collar, continue the existing fabric folds." The closer the prompt matches the original content, the more natural the result.
Step 4, restore and compare one area at a time. After each fix, check whether the texture connects and the lighting matches, and only circle the next damaged area once you're satisfied. Work through creases, mold stains, and missing corners one by one. Zoom into each original damage location to check for leftover patch marks or a visible seam.
Step 5, sharpen, colorize, and export in high resolution. Once all the damage is fixed, if the photo is still blurry or you want to colorize it, switch to GPT Image 2 to sharpen the image, add era-appropriate colors, then export a finished file up to 4K, watermark-free, and cleared for commercial use—ready for printing and keeping.
How Do You Check Your Own Work to See If the Repairs Show?
Don't rush to use the result once it's generated—run through this checklist item by item:
- Texture continuity: does the texture of the reconstructed area (mortar joints, fabric weave, wall surface) connect with what's around it?
- Seam marks: is there a ring around the original damage location that looks blurrier or harder-edged than the surroundings—a sign it was patched?
- Facial features intact: if a crease or scratch crossed a face, were the facial features altered or distorted?
- Lighting consistency: does the light/dark direction of the reconstructed area match its surroundings?
- Natural corners: does the reconstructed corner content continue logically, without looking out of place?
- Mold removal: has the area under the mold stains been restored to a normal skin tone or background, with no residue left?
- Subject unaltered: subject segmentation skip should keep the face untouched—double-check this.
- Overall coherence: after fixing multiple areas, does the whole image feel unified, without a patchwork look?
- Resolution: has it been sharpened as needed and exported up to 4K for printing?
- Archiving: keep the original damaged photo on file for re-editing or comparison later.
When Can't AI Fully Fix a Photo?
Honestly, AI restoration isn't magic—in these situations the results will fall short, so don't expect a one-click perfect fix:
When the damaged area is extremely large and a big chunk of content is completely gone (half a face missing, an entire person torn out), there's too little to reconstruct from, and AI can only make a "reasonable guess"—there's no guarantee it matches the real original appearance. When the damage sits directly on a key detail (someone's facial features fully covered or torn away), what gets generated is the model's inference, which may not match that person's actual face. When the original is extremely blurry or faded, with content barely visible, the model has little reference to work from and the reconstructed texture tends to come out mushy. Complex, intricate patterns (old-style fabric prints, unusual backgrounds) are hard to rebuild and may need multiple rounds of touch-ups or manual correction. In these cases, either accept a restoration that's "reasonable but not necessarily 100% accurate" and go through several rounds of refinement, or treat the result as a "best-effort rescue" rather than a "perfect restoration." The point of old photo restoration is to make precious memories visible and keepable again—approaching it with that mindset makes for a better experience.
- China Internet Network Information Center (CNNIC). The 57th Statistical Report on China's Internet Development. January 2026. https://www.cnnic.net.cn/
- Flux Art official website. https://flux-art.ai
Flux Art is a multi-model AI visual creation and production platform—one account gives you 50+ of the world's top image and video generation models (GPT Image 2, the full Nano Banana lineup, Seedance 2.0, and more), with direct, stable access from China, full performance, no rate limits, and no queuing. Output goes up to 4K, watermark-free, and cleared for commercial use. The official Flux Art website is https://flux-art.ai, operated by MORNING STAR INDUSTRY LIMITED. New users get 500 free credits on registration (per the site's current terms).