
Product and Marketing @previewio | Storyteller | Director | Pursuer of Weird
Joined November 2021
- Tweets4.5K
- Following636
- Followers2.1K
- Likes15.8K
Pinned Tweet
What If?
This took me 1 day to write
4 hours to shoot with an old friend
2 days for the AI and editing
And a ton of fun
Let me know in the comments if you want to see a complete breakdown of how we made this.
Insanely creepy and filled with so many great shots. I loved how often it would take me a moment to realize there was some terrifying tension hidden in the scene.
It’s just a cereal commercial.
PEEPERS | Official Trailer
Film premieres Tuesday 11.17.26
@dreamina_ai #dreaminaoriginals #DreaminaCPP
4
246
The only secret to prompting is knowing this method...
The Method: Prompting is just describing with the clearest most refined detail exactly what you see in your minds eye.
Sometimes what I think helps is writing down all the bullet points or key visual elements of whatever you are trying to create.
If you say "give me cat" the model gonna give you cat. And its gonna be a little different and a little the same every time.
But if you say, "Green cat, fur like a shag carpet, hair on its head in the style of a pompadore haircut, a white jagged stripe down its back. Each paw has long gnarled claws fully extended. An awkwardly mischievous expression on its face, small almost closed blue eyes swirling like the cosmos and a single snaggle tooth on the upper right lip. Crouched ready to pounce in the shadows of a neon lit market in the belly of a dystopian city furnished like a Victorian circus."
Youre gonna get something super specific. Which then allows you to very easily pinpoint what's going wrong and not adhering and then you can surgically adjust your text to get your desired output.
To get whats in your brain visualized.
Learn to clearly describe exactly what you want and the rest will follow.
1
195
Great breakdown of a unique concept.
I think working with gen ai in this way is incredibly important.
Creativity and expansion of the art is what pushes this space forward much faster than only trying to replicate realism and pure cinema.
(I love pure cinema)
6
320
AI Video tricks that are dead:
- Start and End Frames
- JSON Prompts
- Baking an entire shot into a single image and processing each one by one praying for consistency
What method is next on the chopping block?
Incredible work by @julienvallee
Such high quality with so much talent and clear vision in one short commercial.
This is what years of experience looks like
our opening film for @mondaydotcom ELEVATE 2026 is here!
we teamed up with @invideoOfficial to see how far we could push Agent 2. turns out, pretty far.
huge thanks to Dan Brooks & Leif Huron for somehow letting us get away with this.
music by SonBon studio.
8
493
This is blowing my mind!
The guy behind me is not real.
I'm working on a hybrid production for Preview and this is the first time I've seen an AI actor integrated with footage of my actual performance.
Its so seamless I can't believe it.
The Prompt:
Edit the complete original mirrored video [OG Video Ref] at its exact source duration. The source video controls Kyle's identity, anatomy, hair, entire performance, camera angle, movement, tight framing, all spoken words, original voice, lip synchronization, repetitions, pauses and resets. Preserve these without retiming. Do not mirror again.
CHARACTER: Kyle remains screen-right looking screen-left. Replace only his clothing using [character sheet wardrobe ref] as wardrobe authority, not anatomy or pose.
SUPPORTING CAST: Add Assistant [assistant sheet ref] in the open screen-left space beside Kyle, just slightly behind him, with separate silhouettes. Place him close to Kyle's depth plane: almost equal apparent head size, slightly wider upper-chest-to-head framing rather than a distant waist-up figure. Keep both faces sharp with sufficient depth of field. Assistant watches the offscreen pirate screen-left with a mildly furrowed attentive brow and slow agreeing nod, receptive rather than upset.
LOCATION — INTERIOR SOUNDSTAGE: [location ref] supplies reverse-stage geometry and curtain walls only, not its illumination or wide framing. The pair stand near the edge of the offscreen white cyc with black curtains several meters behind them. Cyc and pirate stay offscreen in front of them.
KEY PROPS: Assistant's approved tablet is raised into the bottom of his close framing to allow visible fingertip contact, without covering his face. For each dialogue cue he looks down, raises it, taps a few entries, then fully lowers it below frame and looks back screen-left. Its faint unreadable screen glow is weaker than ambient facial illumination.
VISUAL STYLE: Preserve source camera and Kyle's framing; position Assistant physically nearer rather than zooming or widening Kyle. Natural photographic detail, crisp readable faces.
GLOBAL LIGHT & COLOR: Broad diffuse neutral frontal bounce from the offscreen illuminated cyc provides normal readable brightness across both faces and clothing. Soft light wraps both cheeks and eye sockets, lifting hard cast cheek shadows with low-contrast gentle modeling and retained skin texture. Foreground people are brighter than the dim background. Background overhead fixture is dimmed and distant curtains stay dark. Neither man is under a hard spotlight. Stable exposure and neutral skin tone, not dramatic underexposure.
SOUND DESIGN THROUGHLINE: Preserve the complete original audio, all repeated lines, words and pauses. Assistant remains silent.
NEGATIVE / EXCLUSIONS: No new Kyle gestures, altered dialogue, additional people, visible pirate, cuts, camera repositioning, readable tablet interface, added music or speech.
Shot 1 [00:00–source end]: One uninterrupted edit. Assistant choreography below uses approximate timing from the original; synchronize to actual source speech without changing it.
[00:00–source end] On 'A little less scary and intense...', The assistant performs the full action: agreeing nod, downward glance, tablet raised into frame, several visible taps, then tablet lowered completely as Kyle ends 'something he can't find' around 00:08. Return attention screen-left for any remaining source tail.
CONSTRAINTS: Exact full source duration, no shortening or extending. Preserve Kyle's entire recorded performance and soundtrack. One continuous shot, zero cuts, stable identities, clothing and hand anatomy. No text, logos or subtitles.
If your AI dialogue scenes come out stiff or rushed, look at your prompt. Most of them only say who talks and when.
The model takes what you write as literally as it can and fills in the blanks with whatever it wants. So you get flat voices, no breath, and two people reading lines at each other.
After a lot of trial and error directing dialogue in Seedance 2.5, these 6 things made the biggest difference
I also made a full walkthrough showing you the exact prompts and specific examples on YouTube - Link in Comments...
1. Write like a novelist - describe every little detail.
2. Give every shot an emotional objective, not just an emotion. "He's sorry" gets you a sad face. "He offers an apology while watching for permission to relax" gets you a performance.
3.Write the actions between the lines. A small inhale before he speaks, a listening face after. That's how you build pacing, tension and release.
4. List the timing of moments in seconds and frames. "Hold his scowl from 00:01.02 until 00:02.00 then at 00:02.00 his eyes look down, he twitches his nose and sighs for a half a second, at 00:03.00 he looks back up with an intense slowness and stares straight into camera with a devilish grin that grows suspiciously wide at the corners of his mouth." gives you a real beat instead of the next actions.
5. Direct the delivery. Say which word gets the weight and whether the line trails off or lands hard.
6. Give each character a reason to be in the scene. Blocking and eyeline alone play stiff. Write what they want or how they feel about the other person in any given moment.
Direct from inside your image instead of text
Annotate is superior to in-paint
Do anything, change enything
Stop writing paragraphs to describe your scenes to the model.
Draw the action.
Annotate mode in Preview lets you
- Freeform draw
- Write notes
- Add images
- Draw shapes
- Edit your annotations
Then generate the video from the image for ultimate control of your scene.
Try it today at preview.io
3
354
Stop rewriting your whole prompt every time Seedance 2.5 rejects it.
Sometimes it's one line the model is misreading: a sound effect, a music cue, a prop, a character, an action.
When nothing you asked for should be restricted, but it gets blocked anyway.
This is what I do:
Go back through the prompt and find the element that's likely causing the rejection.
Cut just that element and run the rest.
Once that version is approved and generated, run the video through edit mode and ask for the piece you cut.
This is for rejections that don't make sense. If you're asking for something NSFW, or something thats actually an infringement, it usually won't help.
4
623
Filmmaking isn't a solo sport. It's a collaborative artform.
Gen AI is no different. Anyone who's tried to make long form content knows - its easier with a team.
Working with a team means managing people and assets becomes just as critical as the creative.
I love Preview because we've built it to embody this philosophy. The platform really is built for collaboration
For filmmakers who know what they want, but need help keeping everything together.
You're still in creative control. But Preview makes sure you get where you're going without the headache.
Film productions can look like total chaos to an outsider.
The key to success is keeping your crew coordinated from start to finish.
Until now, collaborative gen AI productions have been scattered messages and assets across a handful of platforms.
Preview is the solution to this management nightmare.
Everything your productions needs, in one place, with a complete recorded history of all decisions made.
Every asset can be shared, managed and manipulated by whoever you choose on your crew.
Learn more and Sign up at - Preview.io
11
588
Do you use a style reference image or describe the look in the prompt?
My checklist:
One generation, or a couple that barely change: text only is fine. It's the fastest way to work.
A few simple generations: still text, but be more specific. Name the camera, the lens, the lighting and the film look.
A long project across many shots: put the style on your reference assets first, then generate the video from them. Maybe restate style in the video prompt as well.
Anytime the look has to match across a lot of generations, use reference assets.
Example of a strong style prompt - video result in comments:
VISUAL STYLE: Photorealistic Super 16mm film-camera photography, natural 24 fps cadence and 180-degree shutter motion blur. Raw indie texture: coarse irregular organic photochemical grain that changes frame to frame, delicate gate weave, lower contrast, gently lifted textured shadows, restrained saturation, soft nonlinear highlight shoulder and subtle warm halation. Real skin pores and cloth weave, no artificial sharpening or skin smoothing. Spherical 32mm cinema prime on Super 16 at approximately T1.5, fixed focal length. Natural spherical perspective, round bokeh highlights, mildly softened frame edges and subtle spherical aberration in backlight. Shallow depth of field keeps city and river soft while faces remain legible on their shared focus plane. Soft circular ghosts and restrained warm veiling flare, not horizontal anamorphic streaks. Focus pull follows faces during physical approach and lands on both profiles and lips; a tiny human correction is acceptable, never focus hunting. Background falls further out of focus as camera approaches. Muted burgundy, oatmeal wool, faded olive cotton, cream tee, amber skin highlights and blue-gray river shade.
1
7
425
Stop gambling with production design in your prompts.
Here's how I made real looking movie posters with consistent text on the walls in my commercial.
I wanted to make posters for the actual movies that were in the commercial using images from the movies and their titles as text on the posters.
I started with the location and mapped out the positions with blank white posters on the walls.
I told the Preview agent to pull iconic still frames from each of the movies in the commercial we were making and then create titles for each of those movies.
The agent pulled the stills, came up with titles and then used the stills as references to make each poster.
Then I used those posters are references and placed them into the location image. Then used that final image as my location reference, plus all the posters as references as well so that the text would stay consistent when animated with seedance 2.5
Poster prompts and reference flow in the replies below ⬇️
5
10
622
Prompt: Create a portrait urban chase thriller poster titled exactly 'NO CLEAN EXIT'. Use {take} for the weathered charcoal vintage muscle coupe, dense city architecture and warm amber versus cool steel photographic palette, not the source framing. Original aggressive low front-three-quarter composition: coupe cuts diagonally across foreground, towering city canyon bends into deep perspective behind, two pursuing police cars reduced to sharp red-blue lights. Driver stays shadowed behind windshield. Tire smoke and streaked street reflections communicate desperate momentum; physical car geometry remains crisp. Dense 35mm grain, controlled horizontal flare and dramatic late-day sunlight across dented metal. Huge slanted condensed cream title stacked above car with clear letterforms, small tagline 'Every turn has a price.' Sophisticated gritty action key art, no fireball or gore. Flat edge-to-edge poster, no frame, branding or billing clutter.
1
1
92
I tried lighting AI images with real cinematography language.
10 lighting setups, 3 models.
Only Seedream v5 Pro followed the light: 7 of 10.
The Prompts, and where Seedream broke 👇
HOW TO WRITE STRONG LIGHTING PROMPTS
Say where the light is, how big it is, what color it is and what it is doing.
"A large soft 3200K key sits 45° off axis frame-right, slightly above eye level. The camera is on the shadow side, so a Rembrandt triangle falls on her frame-left cheek." That gave a directional soft key.
Choose your fill on purpose, because it changes the shot. Black negative fill gave a deeper, harder shadow side: "Black negative fill frame-left deepens the shadow side."
A white bounce with a stated strength gave a bright, open shot: "A white bounce frame-left lifts the shadows to about 1.5 stops under the key." Same key, clearly different images.
Use looks the model already knows. "An on-camera 5600K ring flash sits around the lens, on axis" gave flat, frontal light with a hard shadow edge on the leather behind her. "Hard cool 5600K streetlight comes through venetian blinds" put stripes across her, the booth and the walls.
Colored and visible light works too. "Spilling teal and magenta fill onto the shadow side of her face, with a magenta edge on her hair from a neon sign deep in the background" gave the color contrast and the neon.
A single small hard Fresnel at 60° with no fill gave a noir look. A bare bulb hanging above the table gave a pool of light on the marble.
If you use more than one light, say which one is the key and how strong the others are compared to it. More on this below...
1
109
WHERE SEEDREAM BROKE AND WHY
The Main Takeaway - Name the key, then give every other light and the shadow side relative to it. For example, "the key is the brightest source; the backlight sits 3 stops under it."
All three misses came from my prompt, not the model.
Setup 6: Soft key plus a hard backlight through haze: the beams rendered, but the hard backlight overpowered the soft key. My prompt never said which light was brighter, so the model picked the more dramatic one.
Setup 3: A table lamp as the key with negative fill on the far side: the lamp keyed her face from the left, but the negative fill barely showed. I never said how much darker the far side should be than the key, so there was no ratio to follow.
Setup 4: "lit only by practicals" the lamp did not light her. A fill from off screen became the real key, so the lamp side of her face ended up darker than the other side. I never named which light was the key. The difference between setups 3 and 4 is one sentence: naming the key.
LIGHTING DIAGRAMS DON'T HELP.
I added a top-down diagram of each setup. It didn't improve the light. It changed where the lamps and framing landed, and in 5 of 20 images it built the flags, stands, a softbox or a bounce board into the shot. The 20 text-only images had none.
52










