EP 049 · 70 MIN NANO BANANA MADE ONE IMAGE A STORYBOARD
Midjourney Token Bashing, Nano Banana, and Weavy with Marco Isle
The short version
Episode 49 pairs Rory and Drew with creative polymath Marco Isle to pressure-test Midjourney prompting and Google's Nano Banana image model. Marco demonstrates token bashing, fusing two prompt words into one, and his custom GPT for extracting reusable tokens from reference images, while Rory shows a Weavy-automated workflow that turns a single Midjourney product photo into a full multi-angle, multi-variant shoot via Nano Banana. They also weigh Nano Banana against Flux, Runway, ChatGPT and Veo 3 for consistency.
Key takeaways
- 01
Token bashing fuses two prompt words into one novel term (like 'illumination' + 'provocation'): Marco says it unlocks results a structured prompt wouldn't find.
- 02
Shoot a subject from front, side and rear three-quarter angles first: that reference set lets Nano Banana generate any new angle without morphing.
- 03
Rory automates the whole shot list in Weavy: drop in one reference image and the flow outputs front, side and rear variants with zero manual prompting.
- 04
Rory calls Flux 'flat' with no depth, and says ChatGPT image gen is stuck at roughly 3:2 crops; both lose to Nano Banana on control.
- 05
To force a non-square aspect ratio in Nano Banana, upload an empty frame of that ratio first — just asking it to 'change to 16:9' fails.
- 06
Marco's Book of Jonah short took about six months and roughly 4,000-5,000 images despite strong per-shot consistency, the hosts note.
- 07
Aspect ratio is a hidden variable: Marco and Rory both say the same prompt produces very different images depending only on the aspect ratio chosen.
Receipts · said on the record
“And then I was playing with like this experimental thing where you like fusing some of the words together because obviously an LLM doesn't look at text how humans look at it.”
“Flux, to me, is flat. I don't like Flux because it looks flat. There's no depth.”
“My default is just like mid-journey for exploration because nothing even comes close to it … Never will. It's just game over.”
“It still is the only tool that, to me, that allows you to be that creative and not have to force control … It is about just like, let me smash two random words together, and maybe this will make something cool.”
“Yeah, it's like, am I gonna get it? Am I gonna get it? Nope, nope. Okay, I gotta run it again.”
“Diving into not only the work that we create, but then looking at influences and then putting those into the GPT, extracting tokens and then messing with them. It's infinite, right? It's an iterative process that will probably never end.”
Chapters
- 0:00:00 Cold open + guest intro (Marco Isle)
- 0:00:35 Marco’s background: design → UX/UI → fashion
- 0:04:48 Aesthetics vs. mechanics; the cooking/technique mindset
- 0:10:50 “Rome Reimagined” + early Frames testing
- 0:15:59 Token bashing & Marco’s custom GPT flow
- 0:20:59 Favorite tokens: shadowplay & moonlight
- 0:21:32 New token: gobo shadows (why it works)
- 0:22:19 Glitch looks: pixel smear & datamosh
- 0:23:34 Prismatic rainbow light streaks (stacking ideas)
- 0:23:54 Androgynous beauty as a control aesthetic
- 0:25:22 exp sweet-spots (3, 7, 10, 13)
- 0:27:17 “Liturgical drift parade” prompt set
- 0:30:24 Why Aspect Ratio is everything
- 0:34:12 Idea sources: mood boards + Cosmos.so
- 0:38:54 Nano Banana enters the chat (workflow + why it’s different)
- 0:42:15 Case study: Book of Jonah (42 scenes, ~5k images)
- 0:43:03 Video models: Veo-3 notes + start/end frame
- 0:43:33 Weavy & automation; setting guardrails
- 0:50:00 NanoBanana pros/cons; real-world limitations
- 0:52:51 Frames: Kling vs others; aspect-ratio flips
- 0:55:11 Aspect-ratio workaround (whitespace expand)
- 0:57:47 Keyframes consistency test (the car roll)
Questions this episode answers
What is token bashing in Midjourney prompting?
Token bashing, a technique guest Marco Isle demoed, fuses two unrelated prompt words into one novel term (for example blending 'illumination' and 'provocation') to generate imagery that neither word alone would produce. Marco built a custom GPT that extracts and fuses tokens from reference images this way.
How does Nano Banana keep camera angles consistent?
Rory showed that feeding Nano Banana a front three-quarter reference image of a car let it generate accurate side, rear and rear-three-quarter angles with the design details staying close to identical, even though the model never saw the back of the car.
Who is Marco Isle?
Marco Isle is a creative polymath with a background in graphic design, UX/UI design and fashion who has worked in AI image generation for about three and a half years. On Fast Hours he discussed token bashing, his custom prompt-extraction GPT, and his Book of Jonah AI film project.
Which video model do the hosts prefer for start/end-frame keyframes?
Rory says he's using Kling 2.1 for start/end-frame keyframing, and Marco agrees it tests well, because Veo 3 doesn't yet support start/end-frame keyframes at all.
What did Fast Hours think of Flux and ChatGPT image generation compared to Nano Banana?
Rory calls Flux 'flat' with no depth and says ChatGPT image generation is limited to roughly 3:2 crops and can look inconsistent; he ranks Nano Banana above both for consistency and control over edits.
Show notes
Rory and Drew invite creative polymath Marco Isle to do what this show does best: break Midjourney in interesting ways. It’s a token binge from shadow play, moonlight, GOBO shadows, datamosh, pixel smear, prismatic rainbow light streaks, and androgynous beauty plus where --exp actually makes images better (and where it ruins them). They also crack open token bashing (smashing words to birth fresh visuals), show why aspect ratio is half the art, and steal moodboard fuel from Cosmos.so.
Then it gets dangerous. Nano Banana enters the chat, turning one still into dozens of consistent shots with plain-English tweaks. They compare video stacks and frames workflows (Veo-3, Kling, keyframes), talk Weavy automation, and use a real “Book of Jonah” case study (42 scenes, ~5k images) to show how fast this all moves (and where it still breaks). It’s Midjourney’s magic, but with receipts.
Keywords: Midjourney tokens, token bashing, EXP settings, aspect ratio, mood boards, Cosmos.so, Nano Banana, Weavy, frames, keyframes, Veo-3, Kling, AI image generation, Midjourney video, consistency, creative workflow.
