Sora is dead. OpenAI killed the web and app experiences back in April, and the API goes dark on September 24. I mention this first because half the frame-chaining tutorials still floating around from earlier this year use Sora as the example tool, and none of that advice ports over cleanly anymore. If you're building a workflow around AI video right now, you need to be building it on Kling, Veo, or Seedance. Those are just where the working tools live in July 2026.
Which brings me to the thing every client call has been about for the last three weeks: frame chaining. If you've been in any AI filmmaking Discord recently, you've seen the term. It's the difference between a sequence that looks like six unrelated clips stapled together and one that actually reads as a shot.
What frame chaining actually is
The old way of stringing AI clips together was brutal. Generate five clips, cut between them, pray the lighting and motion roughly match. Motion would reset, hair would jump, a camera push would suddenly become a pull for no reason, and there wasn't much you could do about it after the fact.
Last-frame seeding fixes this by generating into a transition instead of cutting between two finished ones. You take the last frame of your existing clip, feed it into an image-to-video model as the starting point, and let the model generate the motion that carries momentum forward. If you're using a first-last-frame tool, you can even supply the frame you want to land on, and the model fills in everything between. VideoGen Extend does this well, holding lighting, motion, and style constant across the join because it's not stitching, it's generating continuity.
I've started building this into client delivery on shoots where we don't have budget for a full reshoot but need an extra six seconds of coverage. It works better than I expected the first time I tried it.
There's now a published n8n workflow that automates the whole loop, and it's the most useful piece of tooling I've picked up this month. It sends your clip to a video-understanding API that describes the last frame, the audio, and the scene. An agent then writes the next scene's prompt using that description, pulls the final frame as a JPG, and feeds it back into the video model to start the next clip exactly where the last one left off. Run it overnight and you wake up to a rough assembly of a two-minute sequence built entirely from chained single clips. I'm testing this on a Web3 summit recap edit right now and it's saving real hours.
The reverse-chain trick nobody's talking about enough
Swap your start and end frames and the model generates the motion backwards. Instant rewind effect, without a single frame of actual reverse footage. There's a tutorial going around using Nano Banana for the stills and Kling 2.1 for the motion that nails this. It looks expensive but costs almost nothing once you know the move.
Use it for reveal edits, memory-flashback beats, or those surreal "everything folds back on itself" transitions that used to require actual VFX time. I used a version of this on a brand spot last month where a product needed to "unbreak" itself on screen. The client assumed we'd shot it practically and reversed it in post. We hadn't shot anything at all.
LTX Studio has pushed this a step further with what they're calling Shot Control. Instead of just a start and end frame, you drop in multiple checkpoint frames inside a single generation: a wide shot, a push-in, a close-up, all defined ahead of time as one continuous generation. The pro tip that's held up every time I've used it: keep visually similar frames close together on the timeline. Big jumps between checkpoints confuse the model and you get warping instead of a clean push.
Avatars are solving a different problem, and that's the point
HeyGen's new Avatar V model is doing something frame chaining isn't really built for: locking identity, not motion. You give it a 15-second recording and it holds your face, your gestures, your presence consistent across a long-form performance. No frame matching needed because there's no join. It's one continuous identity render. The new Look Packs feature takes a single photo and generates a full wardrobe of styled variations from it, which is a real production shortcut if you're building out a talking-head brand presence and don't want to schedule five outfit changes.
They've also folded Seedance 2.0 directly into the platform now, so avatar creators get high-end video generation without leaving HeyGen at all.
My workflow note here, and this one will save you money: write and lock your script first, generate the final narration in ElevenLabs, then upload that finished audio into HeyGen and build the avatar render around it. Don't render first and dub later. Every time I've done it backwards it's cost me a re-render.
What's convincing me this isn't just a toy is who's actually using it. Jia Zhangke made a Chinese New Year short using Seedance 2.0 where he has a conversation with his own AI clone, and both versions of him are generated. He didn't act in it at all. Separately, "Dreams of Violets," directed by the Koosha brothers, screened at Tribeca in June: an entirely AI-generated film reconstructing a real 2026 massacre in Iran from journalistic reports and eyewitness accounts. A major festival deciding that AI generation was the right tool for a serious, documentary-weight subject isn't something I would have predicted this year.
Frame chaining and avatar locking are solving opposite halves of the same problem: how do you make a generated shot feel like it belongs to something bigger than itself. Right now nobody's fully merged the two. Whoever figures out how to chain frames through a locked digital actor. Motion continuity and identity continuity in the same pipeline. Is going to have a real edge in this space for a while.