Someone on my Berlin crew asked me last week which video model we should default to for the next client pitch, and for the first time in about a year I didn't have an instant answer. Sora is basically dead (OpenAI pulled the web and app on April 26, API goes dark September 24). Meta just walked into image generation with a feature that lets strangers edit your public Instagram photos. ByteDance dropped a model this week that makes 30-second native generations look almost boring to type out. It's a strange, crowded moment to be doing this for a living, and I mean that as a compliment.
Meta's Muse is useful, and also a little bit of a mess
Muse Image launched July 7 out of Meta Superintelligence Labs, free inside the Meta AI app, Instagram Stories, and WhatsApp. The prompt-based editing is genuinely good. Erasing photobombers, mocking up landmark selfies. It's the kind of feature I'll reach for on a client shoot when we need a quick placeholder image and don't want to wait on a full Nano Banana Pro pass.
What's getting the attention, deservedly, is that Muse lets you manipulate another person's Instagram images with AI as long as their profile is public. Worth sitting with what "public" actually means in that sentence. It doesn't mean consenting, it means the algorithm decided your account wasn't locked down. Runway and OpenAI both got burned early on likeness and consent issues, and eventually, after enough bad press, they landed on the idea that "technically public" isn't the same as "fair game." Meta shipped this feature anyway, in 2026, well after watching that whole saga play out elsewhere. I don't read that as naivety so much as a deliberate bet that the feature was worth the blowback. Muse Video is reportedly already in development, and given how the image tool launched, I'm not expecting the video version to arrive with meaningfully better guardrails.
Seedance 2.5 is the model actually worth your attention this week
Forget the Meta drama for a second, because the real production news is ByteDance's Seedance 2.5. Early July rollout, up to 30-second native generation, 4K output, up to 50 multimodal references. Compare that to Seedance 2.0's 15-second ceiling and the target is obvious: Kling and Veo's duration limits, directly, on purpose.
I've been doing longer-form brand work where the client wants a continuous 20-second hero shot instead of three stitched 5-second clips pretending to be one. Every extra reference frame and every additional second of native coherence saves an editing pass in post. The fifty multimodal references is the number that actually matters here, more than the 4K headline. It's enough to lock character consistency, wardrobe, and a location plate in one generation call, something that used to mean juggling three separate tools at once.
Meanwhile the rest of the field has settled into a real hierarchy. Kling 3.0 Turbo currently tops the blind-vote leaderboard with an arena score of 2040, ahead of Happy Horse 1.0 at 1785 and Seedance 2.0 Fast at 1752. At $0.11 to $0.14 per second, it's still the value pick I give clients when budget is the constraint. Veo 3.1 is my go-to when the brief needs native audio and realistic motion, priced at $0.40/$0.15/$0.05 depending on tier. I no longer recommend Sora to anyone, not because the model got worse, but because pointing a client toward a discontinued product is a fast way to lose that client.
Midjourney V1 and the digital twin pipeline everyone's copying
Midjourney's video model still won't do straight text-to-video, which I think is the smart call rather than a limitation. Forcing you through an image first keeps you inside Midjourney's actual strength. Aesthetic control. Instead of making it compete on raw motion with Kling. The trick worth knowing: use --raw if the signature Midjourney look is fighting your prompt. It dials back the model's built-in creative flair so your actual text carries more weight in the final frame. For directorial control on a client job where the brand doesn't want "that AI look," it's the single most useful parameter in the toolkit right now.
On the avatar side, there's a pipeline I keep seeing repeated in creator circles, and one I've started using myself for digital twin talking heads:
- Generate a reference image in Higgsfield Soul
- Merge in a real headshot using Flux Kontext or OpenArt's Omni references
- Run it through Enhancor AI if the skin looks too smoothed-over and plastic
- Voice it in ElevenLabs, Alpha v3 model specifically. It holds emotional inflection better than earlier versions
- Animate with HeyGen Avatar 4, and check "more expressive motion" for anything that needs to feel like a person talking rather than a script being read at a camera
That last toggle matters more than people expect. Without it, HeyGen avatars nod like they're on a slight delay. With it, the head movement actually tracks emphasis in the voice line.
The festival circuit is proving something the industry keeps ignoring
Orion, Stefan Flickinger's darkly comic superhero short, took both Best AI Short and Audience Choice at the AI International Film Festival's June program. The Real Real, a two-minute piece from Sebastian Strasser and Rohan Tehrani, swept Most Fun, Most Surprising, and Best Use of AI. Worth noting: two minutes was enough to sweep three categories. The Google Gemini "1 Billion Followers" award required 7 to 10 minute films at 70% AI minimum and still pulled 3,500 entries globally before naming Zoubeir ElJlassi's Lily the winner back in January.
What these festivals keep proving, cycle after cycle, is that runtime and budget were never the real bottleneck. Taste was, and still is. The tools got good enough eighteen months ago; what's changed since is that the people using them have gotten better at knowing what to cut.
I don't think we're heading toward one dominant model anymore, and honestly, good. A year of Sora hype convinced a lot of clients that one tool would solve everything. Now I get to tell them the truth: use Kling for value, Veo for audio-synced realism, Seedance for length, Midjourney for a look nobody else has. It's not a compromise so much as it's what a real toolkit has always looked like. Most clients just haven't needed one until now.