Félix Sanz
@felixsanzm
generative AI, tested · image, video, 3D, audio, LLMs · what actually works · full-stack dev leading DX @runware · 🇪🇸/🇬🇧
474 Following 632 Followers
most AI systems reach for an LLM even when the job is only to choose, score, or route
Jev suggests a different architecture: let generative models generate, and use decision models to decide
Show more
every summer, millions of humans migrate to the all-inclusive resort
a 2 minute nature documentary made with Seedance 2.5
every summer, millions of humans migrate to the all-inclusive resort
a 2 minute nature documentary made with Seedance 2.5
best thing I've pulled out of Seedance 2.5 so far: a full war movie about assembling a bookshelf
somebody had to make Super Pang in real life
always felt bad for the lemmings. seeing them as real people doesn't help
Pong should've been an Olympic sport
one continuous 30s take with Seedance 2.5, shot like a TV final. stay for the last point
The Sims, if it had really happened. sul sul
one 30s drone take made with Seedance 2.5, from a kitchen fire to an uninvited guest at the pool
Age of Empires II, if it had really happened
one continuous 30s drone shot made with Seedance 2.5
ran the four MiniMax H3 tiers on the same 15 second shot
I'd skip the base model, it took over 5 minutes and $1.20
Max Turbo did it in 27 seconds for $0.15, and nothing in the base version looks worth 8 times the price
Fast costs more than Turbo, at half the resolution
Show more
another FLUX Video Edit [fast] example / use case. this time changing words in a dialogue
bye bye reshoots because someone said the wrong word
FLUX Video Edit [fast] changing the season of a shot
you can turn real footage into another style with FLUX Video Edit [fast] and the performance stays exactly as shot
something I like about Gemini Omni Flash 1.1 is the video extension feature, works so nicely!
something unique about Meta Muse: when you ask for a chart, it writes and runs code to plot your data before it draws anything
so it isn't guessing bar heights and axes like other image models do. top one without it, bottom one with it
Show more
Recraft V4 Styles builds a reusable style from a few reference images. no LoRA, no training
one style, four different scenes here, all the same look. you reuse a styleId to stay consistent
works with SVG outputs too
Show more
this is not original NASA footage. it's not restored or colorized archival film. every frame was generated by Wan 3.0 from a single web page
I gave it one thing: the Apollo 11 mission page on no script, no clips, no images. it read the page, extracted the mission's beats and filmed them
a URL in, a video out
Show more
this one is wild. Wan 3.0 turns a document into a video
I gave it a plain markdown recipe file for beef tacos, nothing else. it read the ingredients and the numbered steps, and generated an overhead cooking video that follows the recipe in order, pulling the exact spices from the source
it takes Word, PowerPoint, spreadsheets, Markdown, and public webpage URLs too. the scripting and storyboarding step just disappears
Show more
