Iconic Film from 1896 Restored with Artificial Intelligence: Watch an AI-Upscaled Version of the Lumière Brothers’ The Arrival of a Train at La Ciotat Station

Machine learn­ing keeps, well, learn­ing in leaps and bounds, and at Open Cul­ture we have watched devel­op­ments with a fas­ci­nat­ed, some­time wary eye. This lat­est advance checks off a lot of Open Cul­ture box­es: trav­el­ing back in time through the pow­er of film; home­grown inge­nu­ity; and film his­to­ry.

YouTu­ber Denis Shiryaev took the lat­est advances in AI tech and turned them onto one of the ear­li­est works of film: The Arrival of a Train at La Cio­tat Sta­tion, shot by the Lumière Broth­ers in 1896. There are plen­ty of urban leg­ends around this 50 sec­ond short: that it was the first ever Lumière film (it wasn’t, they had a selec­tion of pre­vi­ous shorts); and that audi­ences were ter­ri­fied, think­ing the train would hit them (they were amazed, no doubt, but they weren’t that naive).

You might want to watch the orig­i­nal below before watch­ing Shiryaev’s 4K upscal­ing and AI “smoothed” ver­sion to get a sense of the mar­vel at the top of the post.

What we are see­ing is not a tra­di­tion­al “restora­tion,” how­ev­er. Instead, Shiryaev is using a com­mer­cial image-edit­ing soft­ware called Gigapix­el AI. (If you have the pro­cess­ing pow­er, you can try it out). The orig­i­nal film was not shot at 60-frames-per-sec­ond. Instead, neur­al net­works are look­ing at the orig­i­nal frames and “fill­ing in” the data in between, cre­at­ing what you can see is a more nat­u­ral­is­tic effect. Peo­ple on and off the train move like they do in real life. It looks like it was shot yes­ter­day.

Now, this isn’t per­fect. There are a lot of arti­facts, squooshy, mor­ph­ing moments where the neur­al net­work can’t fig­ure things out. But hey, this is just one guy on his com­put­er. It’s an exper­i­ment. The com­put­er code will get bet­ter.

The Gigapix­el AI was devel­oped by Topaz Labs orig­i­nal­ly to help pho­tog­ra­phers upscale their pics by 600 per­cent with­out los­ing detail. It didn’t take long to apply this to video, but be warned, it can take hours of pro­cess­ing pow­er to ren­der a cou­ple of sec­onds. Still it hasn’t stopped peo­ple from exper­i­ment­ing, even with sim­i­lar neur­al net­work pro­grams:

Here’s a clip from Nirvana’s “Heart Shaped Box” video upscaled to 4K with Gigapix­el AI:

User AkN upscaled A‑Bomb footage from the 1950s:

Some clips from Home Alone:

You get the idea. As with any tech­nol­o­gy, there are also some hor­rif­ic exam­ples out there too where it just does not work. But I have a feel­ing that Shiryaev’s first dive into film his­to­ry is not going to his, or the internet’s, last.

Relat­ed Con­tent:

Dra­mat­ic Col­or Footage Shows a Bombed-Out Berlin a Month After Germany’s WWII Defeat (1945)

Pris­tine Footage Lets You Revis­it Life in Paris in the 1890s: Watch Footage Shot by the Lumière Broth­ers

Immac­u­late­ly Restored Film Lets You Revis­it Life in New York City in 1911

Ted Mills is a free­lance writer on the arts who cur­rent­ly hosts the artist inter­view-based FunkZone Pod­cast and is the pro­duc­er of KCR­W’s Curi­ous Coast. You can also fol­low him on Twit­ter at @tedmills, read his oth­er arts writ­ing at tedmills.com and/or watch his films here.