Restoring old footage with AI, the full breakdown I tested two things for this and both gave good results, Higgsfield AI 🧩 Genjutsu and MiniMax Design (H3) MiniMax H3. Steps and prompts below 👇

1. Start with the first frame Take the first frame of the original clip and run it through GPT Image 2 in two steps, restore first, then upscale. • Restore prompt: Ultra-realistic restoration and recreation of the uploaded old vintage photograph. Preserve the exact same person and original facial identity with 99% likeness — same facial structure, eyes, eyebrows, nose, lips, jawline, hairstyle, expression, age, skin tone, pose, body proportions, and camera angle. Do not alter, beautify, redesign, or replace the person. Repair scratches, dust, stains, fading, blur, discoloration, noise, creases, and damaged details while keeping the photograph authentic.Transform it into a clean, sharp, high-quality modern digital portrait with naturally updated, vibrant colors, realistic skin texture, fine facial details, balanced contrast, and soft natural lighting. Upgrade the clothing and background into a clean, refined modern aesthetic while preserving the original outfit style, composition, pose, expression, and emotional character of the photograph. Keep all restored details believable and historically respectful. No artificial-looking skin, face distortion, identity drift, excessive smoothing, exaggerated colors, or unrealistic features. Photorealistic, professionally restored, ultra-detailed, high resolution. • Upscale prompt: Upscale this exact image @[Image 1] to 4K resolution. Preserve the original image completely — same face, same identity, same facial structure, same expression, same gaze direction, same head angle, same skin texture and wrinkles, same lighting, same shadows, same colours, same clothing, same background, same composition and framing. Change nothing. Only increase resolution and recover fine detail: sharpen existing skin pores, stubble and beard hairs, eyelashes, fabric weave and the edges already present in the image. No stylization, no beautification, no smoothing, no skin retouching, no wrinkle removal, no colour grading, no relighting, no added detail that is not already in the original, no change to face shape or proportions, no change to expression.

2. Feed both into the model Give the original clip and your restored, upscaled first frame to Genjutsu or MiniMax H3 with this prompt. The {} parts change with every clip. Paste your frame into Claude or GPT and have it fill them in 👇 • Prompt: MOTION ONLY @[Video 1](video_1) is a motion plate. Use it for pose, timing, lips, hands, camera and framing. Do not copy its colour. Do not copy its grey look. It has zero chroma. COLOUR SOURCE All colour comes from @[Image 1](image_1). Copy that image's colours onto every frame of the performance. RESULT Full colour in frame 1, frame 2, and every frame after that. Skin: {skin tone, lip colour} Hair: {hair colour, length, texture} {Clothing item 1}: {colour, fabric, detail} {Clothing item 2}: {colour, fabric, detail} {Prop or instrument, if any}: {colour and materials} Background: {every coloured element in the scene, listed} HARD RULE If the model starts to go grey, push back to @[Image 1](image_1). No monochrome. No sepia. No half-colour. No flicker between colour and grey. Not a restored kinescope. A colour studio recording. Same {man/woman}, same face, same hair, same {clothing}, same {prop}, same gestures, same lips, same camera, same timing. No subtitles, captions, watermarks, logos, burnt-in text or lettering anywhere in the frame. Clean sharp digital image. No grain, dust, scratches or flicker.

• Comparison 👇 Genjutsu is more accurate on motion control and timing, H3 drops the timing on some frames. H3 wins on cost, and on copyright: Genjutsu filters old footage of well known people. Genjutsu wins on realism and detail, and it goes up to 30 seconds in one clip while Hailuo stops at 15.

@higgsfield_ai @Hailuo_AI This isn’t “restoration”; it’s speculation. A dangerous approach to archive footage.

You're right, this was my first experiment using the new tools on old footage. The thing with archive footage is that it varies a lot clip to clip. Some are good enough quality that colorizing alone does the job. But in others, details and objects in the environment are barely readable. What approach would you take there, if not regenerating?

@higgsfield_ai @Hailuo_AI Fake colour wrecks archive footage. It should be avoided at all costs. It's presumptious guesswork that doesn't conform to the light conditions present when the film was shot. My approach would be to preserve the footage in as close a condition to when it was found. 1/2

@higgsfield_ai @Hailuo_AI These are historical documents and should be archived correctly. "enhancing" them sets a dangerous precedent if planning to present them as genuine: their documentary value is lost and they become simulacra. 2/2

@higgsfield_ai @Hailuo_AI Someone else I saw did the Bruce Lee one but idk how they did it.

This is a really solid result! Looks like a combination of Topaz and another motion transfer tool. For better consistency, upscaling multiple keyframes usually yields way better results. I wanted to do that, but Genjutsu kept losing color consistency for some reason, so I stuck with a single frame. The downside of single-frame is that as the clip moves away from frame 1, unwanted details start popping up. But when you lock multiple frames, it acts as a proper motion control between them. Definitely improvable with more testing!

@higgsfield_ai @Hailuo_AI This is not restoration, you're making a reenactment. Restoration would be restoring the information that was actually captured in the frame but this is creating all new information that was never there. Every single frame is significantly different than the original.

@higgsfield_ai @Hailuo_AI Looks great

Thank you🙏

@higgsfield_ai @Hailuo_AI The 2K quality on the MiniMax H3 makes it look so good🔥

@higgsfield_ai @Hailuo_AI Yeah, both look great quality-wise! It was a cool experiment 🙌

@higgsfield_ai @Hailuo_AI Pretty cool fam🤍

@higgsfield_ai @Hailuo_AI Really appreciate it 🙏

@higgsfield_ai @Hailuo_AI Love seeing your experiments Armin, great work

@higgsfield_ai @Hailuo_AI Thank you so much! Glad you liked it.‌

@higgsfield_ai @Hailuo_AI Amazing!

@higgsfield_ai @Hailuo_AI this is bullshit, details are shit

@higgsfield_ai @Hailuo_AI Then do a better one, let's see!

@higgsfield_ai @Hailuo_AI u do better and dont post this crap

@higgsfield_ai @Hailuo_AI Talk is cheap. Show some actual work or just keep scrolling 🤷‍♂️

@higgsfield_ai @Hailuo_AI your work is cheap,it isnt even yours gtfo

@higgsfield_ai @Hailuo_AI Try seven samurai.

@higgsfield_ai @Hailuo_AI Interesting 🤔.. I will try with my family old footages

@higgsfield_ai @Hailuo_AI Old footage looks brand new

@higgsfield_ai @Hailuo_AI This genuinely feels like creativity with a clear artistic purpose.