Here’s Part 2, this time made with Seedance 2.5.
This is the final cut, built from shorter generations, and I’m really happy with how it came together.
My first attempt was one continuous 30-second generation, but it introduced too many mistakes. At 2.5’s price, repeatedly rerolling a long clip just wasn’t practical.
Breaking the sequence into shorter shots gave me much more control over blocking, character placement, vehicles, pacing and the final edit.
The biggest upgrade is the rendering. Skin has more texture and looks far less plastic than it often did in 2.0.
Multimodal referencing is genuinely useful too. I used character, vehicle and environment sheets, but never needed more than seven references at once.
Is it worth the extra cost? Yes, if time matters. I think 2.0 could reach a similar result, but it would take more generations and a lot more work.
A few example prompts in next posts 👇
Made via @dreamina_ai
Second go with @Hailuo_AI#MiniMaxH3
This time I ran gritty Midjourney v8.2 stills through H3 and used a Suno track as reference.
Big takeaway: simpler prompts worked better. H3 understood the intent, stayed close to the original aesthetic, and synced the movement to the track surprisingly well.
There are a few mangled faces and hands in there, but most of that came from imperfections in the source stills rather than H3 breaking the motion. This was just a quick test, so I left them in.
I also used a vérité-style camera block and crash zooms to make the footage feel less controlled and more observed. Handheld movement, slight focus breathing and subtle background drift gave the shots a much more natural feel.
Prompt block for the vérité camera work in the next post 👇
For a while most AI image models gave you the same Shutterstock look, and midjourney was about the only way out.
This year a few models have been pulling us away from that: Luma uni 1.1, Krea, Reve, H3 image, and now I think I'll now add Seedream 5.5 to that list.
I asked for a faded 1950s color negative pushed two stops and I got some grain, helation and stills that feel lived in.
via @capcutapp studio
For a while most AI image models gave you the same Shutterstock look, and midjourney was about the only way out.
This year a few models have been pulling us away from that: Luma uni 1.1, Krea, Reve, H3 image, and now I think I'll now add Seedream 5.5 to that list.
I asked for a faded 1950s color negative pushed two stops and I got some grain, helation and stills that feel lived in.
via @capcutapp studio
Josh has an incredible portfolio of work. There’s a lot of thought and craft that goes into producing AI content at this level. Really valuable to get a glimpse into his process and approach here
Most AI sports content still looks like AI, and fans notice faster than anyone.
One wrong crest, one plastic face, and you'll lose the audience faster than Amorim lost the United dressing room.
Here's how I got real faces for my Morocco film ↓