EP 046 · 73 MIN WE TESTED MIDJOURNEY VIDEO ON DAY ONE

Using Midjourney Video Loops & End Frames: Live Teardown

July 25, 202573 min Rory Flynn & Drew Brucker

The short version

Episode 46 is a live teardown of Midjourney's video update, which shipped loop, 720p and start/end-frame keyframes the night before recording, bugs included. Rory and Drew test keyframe chaining, direction-correction tricks and multi-frame extensions, then pivot to JSON-structured prompting in Veo 3, where Rory iterates roughly ten times to get a gradual truck-to-robot transformation. They also weigh Runway Act 2's motion capture and call Sora's current output 'a flop' by comparison.

Key takeaways

  1. 01

    Midjourney's new video keyframes launched buggy on day one: dragging images into the first frame often failed, and re-rolls sometimes ignored the chosen motion setting.

  2. 02

    If a Midjourney video subject walks the wrong way, just add a direction word like 'walking right' to the prompt; it's surprisingly literal about it.

  3. 03

    You can chain keyframes past two: extend a generated clip, and its last frame becomes a new first frame, letting you build longer multi-scene sequences.

  4. 04

    For complex Veo 3 shots, have ChatGPT write a timed JSON prompt (e.g., 'arms unfold 1-2s, torso folds 2-4s'), then iterate on your output video.

  5. 05

    Rory needed about ten rounds of feeding Veo 3's own output back into ChatGPT to turn a snap-cut robot transformation into a gradual one.

  6. 06

    Midjourney video can't take a dragged style reference; if you want consistent color and style across keyframes, you have to use a numeric style code instead.

  7. 07

    Rory calls Sora 'a flop of a tool' after testing the same JSON video prompts there that worked well in Veo 3.

Receipts · said on the record

“I feel like everyone needs to have a starting point because video is iterative. I don't care what anyone says. Like it's iterative. You're to get 60% of the time, you're to get awesome stuff. 40% of the time it's going to be trash.”
Rory Flynn Rory on why AI video demands iteration, not one-shot luck Play from 1:05:34
“It's been funny trying to throw these JSON prompts. I've just been testing them in Sora also just to watch the thing just crash and burn. It's like, man, what a flop of a tool compared to everything else right now.”
Rory Flynn Rory's blunt verdict on Sora after testing the same JSON prompts Play from 1:11:39
“This is fun social media fodder. I think maybe where your struggle is, if you're looking for product-specific level of detail and no morphing, I don't think we're there.”
Drew Brucker Drew on why AI video is social-ready but not product-precise yet Play from 54:52
“The only problem with what I find with Magnific is that if you have text or graphic or something like on a t-shirt or like logo or whatever, like if you use any sort of level of creativity, zero, one, even, it changes it.”
Rory Flynn Why even minimal creativity settings distort text and logos in Magnific Play from 10:56
“What do you think about that prioritization with V8? Knowing what we know, I guess, versus better things, like improvements around user experience and UI, of like just getting to things faster.”
Drew Brucker Drew questions whether V8 should focus on UX/speed, not just quality Play from 12:53

Chapters

  1. 0:00:00 Intro & accidental nudist beach adventure
  2. 0:02:50 DIY sunscreen & unexpected AI life hacks
  3. 0:07:00 Midjourney video update overview (looping, 720p, start/end frames)
  4. 0:10:20 Upscalers, Magnific precision, and V8 development focus
  5. 0:15:30 Personalization codes & base model quality debate
  6. 0:17:30 Custom GPT for Midjourney knowledge recall
  7. 0:21:10 Mood boards, micro-styles, and avoiding “homogenous AI look”
  8. 0:24:40 Style Explorer, aesthetic preference survey, and upcoming features
  9. 0:27:10 Live first-frame/last-frame keyframe testing
  10. 0:38:30 Loop functionality and extended multi-keyframe workflows
  11. 0:45:40 Iterative prompting lessons and fixing motion quirks
  12. 0:53:30 JSON prompting explained and social-ready video hacks
  13. 0:58:00 Runway Act Two motion capture tests and impressions
  14. 1:07:30 Sloth race cars, Trump in Lord of the Rings & other AI absurdities
  15. 1:09:40 Key takeaways and what’s coming next

Questions this episode answers

What new Midjourney video features launched in this episode?

The episode covers Midjourney video's loop function, 720p resolution, and start/end-frame keyframes, all of which shipped the night before recording and were still glitchy during the live test, including images failing to drag into the first frame.

How do you chain more than two keyframes together in Midjourney video?

Rory and Drew found that extending a generated keyframe clip turns its last frame into a new first frame, letting you keep adding end frames to build a longer, multi-scene sequence rather than being limited to one start/end pair.

What is JSON prompting for AI video?

JSON prompting means having an LLM like ChatGPT write a structured, timed prompt (specific actions at set second-ranges) for a video model like Veo 3, instead of a plain paragraph. Rory used it, refined over about ten iterations, to get a tractor-trailer to gradually transform into a robot rather than snap-cutting.

What did Fast Hours think of Runway Act 2?

Rory says Act 2 didn't feel like a huge leap over Act One, but its real improvement is more natural hand and face motion; in his test, having Jack Nicholson's performance drive his own added realistic gestures, like a high-five, that Act One wouldn't have produced.

How did Sora perform compared to Veo 3 in this episode?

Poorly, according to Rory: he ran the same JSON-structured prompts that worked well in Veo 3 through Sora and called the result 'a flop of a tool compared to everything else right now.'

Show notes

Fast Hours Podcast Ep. 46—Using Midjourney Video Loops & End Frames: Live Teardown (with Drew Brucker and Rory Flynn)

Rory accidentally finds himself on a nudist beach while Drew's making DIY sunscreen with AI.

And if that wasn't crazy enough, this episode is a full live teardown of Midjourney video loops and end frame control—features built for creating cinematic AI video workflows.

Drew and Rory show how to use loops, start/end frames, and extended keyframes to build seamless sequences, plus what to avoid so you don’t burn through credits.

You’ll also learn: Keyframe Extensions – chaining multiple shots for longer, smoother videos JSON Prompting – precision timing and motion control (with live tests) Runway Act Two – motion capture updates and creative comparisons Midjourney Style Explorer & V8 Preview – what’s next for AI-driven video creation

Whether you’re a creative director, designer, marketer, or experimenting with AI video workflows, you’ll get practical prompts, iteration techniques, and creative hacks to level up your Midjourney results.

Watch now to see how these new features work, what to avoid, and how to produce cinematic AI videos faster.

Every word Read the full transcript

If you liked this one