Seedance 2.5, MiniMax H3, FLUX 3, and WAN 3 are now live on Oxen!
Four brand new video releases landed over the past week. Everybody chose to show off and we're very happy they did.
Also, amazing news to report, open source video models are back. We've been praying for times like this and we couldn't be happier they are here.
Seedance 2.5
We all know about Seedance 2.0, it showed up and became the undisputed champion of video gen quality. It feels like it's been years since we saw a model compete with it (it came out about six months ago, which in fairness is years in the AI world). This is ByteDance's attempt at keeping the crown.
Everybody is talking about Christopher Nolan's Odyssey right now, so we gave every model on this list the same brief:
A weathered Greek hero returning home from war, gripping the mast of a wooden sailing ship in a violent storm, bronze breastplate over a red tunic, soaked hair whipping across his face, rain sheeting sideways, waves breaking over the deck, lightning flashing overhead. Cinematic, 35mm, handheld.
You might've realized that the prompt we used for this model is a bit different. Using "Odysseus" in the prompt triggered a copyright error (come on ByteDance, the Odyssey has been in the public domain for about 3000 years). This obviously is ByteDance being cautious after the whole Brad Pitt vs Tom Cruise debacle that took over all of our timelines when Seedance 2.0 released.
Still, this is probably the best looking generation out of the bunch, 5s at 720p for about a dollar, so great as always and pricey as always.
Try Seedance 2.5 Reference to Video
MiniMax H3
Huge news, first time in a long while an open source model is #1 for video editing on Artificial Analysis (and top three for both text-to-video and image-to-video). Worth knowing that the released weights top out at 768p; the 2K path is MiniMax's own and stays on their API, which is what runs here.
Text, images, video clips and audio can all be passed in as input, and you can run this model locally on about 80GB of VRAM.
Most importantly, we can fine-tune this model. This is great news for all of us that have been suffering with Seedance censorship, I have a feeling that this is gonna be the daily driver for a lot of people.
You can fine-tune it without having to worry about code or infrastructure directly on Oxen.ai. We did the heavy lifting, now all you have to do is upload your data.
Odysseus lashed to the mast of a wooden ship in a violent storm, rain sheeting sideways, waves breaking over the deck, lightning flashing across his face as he shouts into the wind. Cinematic, 35mm, handheld.
This 5s generation at 2K cost about 65 cents and it looks pretty great, this is no doubt a great model, but the fact that we can train it puts it way ahead of the competition in our humble opinion.
FLUX 3 Video
Black Forest Labs' first video model, generally available since August 4 after a couple of weeks in early access. You can pass in text, images, or a video to continue from. These multi-input models are becoming the norm, thankfully.
Odysseus lashed to the mast of a wooden ship in a violent storm, rain sheeting sideways, waves breaking over the deck, lightning flashing across his face as he shouts into the wind. Cinematic, 35mm, handheld.
Really good photo-realistic quality. Five seconds at 720p cost about a dollar, and nothing got copyright flagged, which will matter more once you get further down this post.
WAN 3.0
Alibaba's latest warrior entered the arena, and it takes every modality as input too, image, text, video and audio.
From our experiments the output looks very good, though it hasn't turned up in the Artificial Analysis Video Arena yet, which is the leaderboard that ranks on blind human preference rather than vendor benchmarks.
It's also closed, which continues the pattern, Wan 2.2 was the last open flagship model, we miss your open source models Alibaba, please bring them back.
Odysseus lashed to the mast of a wooden ship in a violent storm, rain sheeting sideways, waves breaking over the deck, lightning flashing across his face as he shouts into the wind. Cinematic, 35mm, handheld.
Eight minutes for five seconds at 720p, which is slow, but it cost 50 cents. Half of what FLUX charged for the same clip.
You might also notice that Odysseus is played by Russell Crowe in this particular generation. The copyright and likeness debate isn't getting put to bed any time soon (Maybe Odysseus was related to Maximus Decimus Meridius)
Let us know which one is your favorite, and don't hesitate to tag us in your generations. We'd love to see what you come up with.
If you'd like some credits to test these models out, email us at hello@oxen.ai and we'll send over some creative fuel (aka compute credits 😛).
Have a nice weekend herd!