I tested Seedance 2.0. Wow.
TL;DR
Seedance 2 is the first video model Serio treats as a true editor, not just a generator — he shows it combining up to two images, two videos, and audio in one prompt-driven workflow to swap characters, backgrounds, products, and language while preserving motion.
The practical unlock is ad production and localization at scale — in one demo, a Chinese glasses ad is turned into an English version with a different model while keeping the exact hand motion, wink, camera blur, and timing for cleaner A/B tests across markets.
Prompting Seedance 2 rewards extreme specificity — unlike models where shorter prompts can work better, Serio says Seedance improves when you describe identity preservation, motion, emotion, and even facial muscle movement, often using Claude Opus 4.6 to optimize prompts.
The virtual try-on demo is where Greg visibly gets sold — Seedance puts Serio into a winter outfit filmed in Montreal at -30°, preserves his face convincingly, matches details like boot and pant patterns, and even adds a bear with tracked eye movement and footprints.
AI influencers look close to commercially usable now — Serio shows lip-synced avatars generated from a Nano Banana Pro image that deliver product lines naturally, keep packaging text stable, and create the possibility of cheap, unlimited creator-style ads without shipping products.
Serio still doesn’t think Seedance replaces every model or Adobe — he calls it the default for video generation/editing right now, but says tools like Kling 3 can still win on certain emotional or stylistic looks, while Adobe remains the post-production layer professionals will keep needing.
The Breakdown
Greg opens with the big claim: this is the AI video model people have been waiting for
Greg tees the episode up like a business briefing, not a toy demo: the people who really understand Seedance 2.0 will be able to build AI influencers, faceless media brands, original films, and multilingual ads that actually convert. Serio frames the whole conversation around workflows and monetization — not just “look how cool this is,” but how founders and creatives can build products on top of it.
Multi-input video editing is the real leap
Serio’s first big point is that Seedance 2 can take multiple inputs at once — up to two images, two videos, and audio — then combine them into a single output based on tagged prompt references. His green-screen demo swaps two people and the background in one shot, and Greg immediately fixates on how tightly the model preserves the original motion from pure natural-language instructions.
Why prompting and references matter more here than with other models
When Greg asks about prompt strategy, Serio says he usually starts manually and then uses Claude Opus 4.6 to refine it, because Claude seems to understand vision-model prompting better than GPT in his experience. His bigger point: Seedance wants detail; if you care about identity, motion continuity, or transitions, vague prompts won’t cut it, and strong source references are half the game because they communicate taste better than text alone.
The virtual try-on demo is the first jaw-drop moment
Serio shows a clip of himself in freezing Montreal weather, then uses Seedance to put himself into a different outfit and add a bear walking by. Greg’s reaction is basically disbelief: the face stays consistent, the clothing pattern and boots match the reference, and tiny touches like the bear’s footprints and Serio’s eye tracking make it feel less like a gimmick and more like a real commercial workflow.
Translation plus character replacement makes global ads feel trivial
Next comes a Chinese glasses promo that gets reworked for a US audience by swapping in a new reference model and translating the spoken content to English. What makes the demo stick is that the exact body language survives — same wink, same hand-to-glasses motion, same focus blur — so Greg starts talking about “content in a hundred languages” and ultra-clean demographic A/B testing.
Product placement and packaging edits now happen inside the video itself
Serio then replaces a blank package inside a rendered ad with a branded product image, keeping logo consistency and the yellow background intact. He compares it to what Nano Banana did for image editing, except now it works in motion, which turns generic stock-like templates into customizable ad assets instead of one-off renders.
Video extension and AI influencers push this from cool to operational
Two more use cases round it out: first, extending a short clip into a longer scene while preserving continuity, which Greg says has been a personal pain point in ad making. Then Serio shows AI influencer lip-syncing, explaining that realism comes from prompting muscle movements and emotional transitions rather than saying “she looks sad,” and the sample ad lands hard enough that Greg says he got goosebumps.
Seedance is the new default, but not the only model that matters
In the closing stretch, Serio calls Seedance the best video model currently available, especially for editing, while noting limits like 720p for now and acknowledging that other models can still win on speed, price, or a specific visual feel. On Adobe, he’s surprisingly measured: generative tools may reshape the front end, but he thinks serious creatives will still need post-production control, meaning Adobe’s long-term relevance likely depends more on agentic editing than on being the place where content gets generated first.