Product and Marketing @previewio | Storyteller | Director | Pursuer of Weird
Joined November 2021
- Tweets4.4K
- Following627
- Followers2K
- Likes15.6K
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.
Kyle Salazar retweeted
This is the easiest way to relight your scenes
From Preview to @Photoshop in one click
Achieve perfect lighting in every shot
Your work stays intact with two way sync
You run the creative, we handle the connection
Kyle Salazar retweeted
Seedance 2.5 draft mode is now in Preview
Generate super fast with 480p drafts
Then upscale to 1080p with a single click
Your original video stays exactly the same
Same prompt same seed same everything
This is a great way to test prompts before committing to expensive 1080p outputs
A breakdown of the entire production process for the Hybrid "What If" video.
Talk through a bit of the shoot, managing footage, the selects process, gen ai choices
And of course how it all came together in Preview
Kyle Salazar retweeted
“The Miracle” Premieres Now:
I gave this film everything I had. Pushed myself creatively and technically, experimented a ton with new tools + models and at times got stuck, started over and wondered if I could pull it off. Excited + nervous to finally share it with you. Enjoy!
No more "Damn how do i fix this video prompt?"
Preview agent can now watch your video.
Tell you where it went wrong.
Suggest a fix and make the change.
Then iterate until it's right.
Meanwhile, you get to work on the fun stuff.
@GoogleDeepMind just launched a new API that allows you to analyze videos.
What I realized after adding that capability to our agent in @previewio , was that the new @fal H3 Model and this together unlock "loops" for video generation.
Generate > Review > Generate
Result:
When you combine GrokBots x Preview CLI to make a commercial...
All I prompted was:
"this one wants to be like a 1950s informational video. Just a guy who works at the company PREVIEW and he's got computers and he's explaining to the camera what PREVIEW is in that old 1950s voice."
Slopmaxxing my prompt skills for the sake of automation experiments.
🤖 Made with AI
Tested Minimax H3 max all weekend. You can use it in Preview now - its cheap and fast.
Best prompt format is in the comments 👇
Here's what you need to know:
- H3 feels like a draft model, it brings a new way to conceptualize and brainstorm shots, characters, locations through video instead of images
- Don't expect to get great quality from t2v especially if there's a lot of action or motion in the shot.
- JSON prompting works best. I'm not kidding. Normal text works well too, but json seems more consistent in its adherence.
- The quality is only as good as your start and end frames.
- You can pack a lot into a little, but the more that's happening the worse the quality becomes.
- Wide shots struggle with character consistency.
- The emotion and voice acting from characters is impressive.
This model really shines on simple things like creating character turnarounds, or a near static frames from a strong reference - see what I mean below.
🤖 Made with AI
JSON PROMPT TEMPLATE
For a single continuous shot, omit the "timeline" array entirely and describe the action in "scene.action". For multiple beats or cuts, use "timeline" and leave "scene.action" as a brief overview. Keep every key present; use "none" or an empty array rather than deleting a key.
{
"visual_direction_only": true,
"narration_instruction": "Visual direction only, no narration read aloud.",
"look": {
"medium": "[format, film stock/grade word, era]",
"presentation": "[how/where this is viewed or framed, or 'direct']",
"lighting_quality": "[quality/color/hardness of light]"
},
"camera": {
"shot_type": "[close-up, wide, medium, etc]",
"movement": "[static, push-in, handheld, pan, etc]",
"movement_change": "[if movement shifts mid-shot, describe when/how, or 'none']"
},
"scene": {
"overview": "[who is in frame, setting, what is happening overall]",
"action": "[direct action description for a single continuous shot; brief overview if using timeline instead]"
},
"timeline": [
{
"timecode": "[0-Xs]",
"camera": "[shot/movement for this beat]",
"action": "[what happens in this beat]",
"dialogue": "[line spoken in this beat, or 'none']"
},
{
"timecode": "[X-Ys]",
"camera": "[shot/movement for this beat]",
"action": "[what happens in this beat]",
"dialogue": "[line spoken in this beat, or 'none']"
}
],
"performance": {
"pacing": "[slow/fast, deliberate/urgent, or 'none']",
"breathing": "[timing/rhythm/breath notes, or 'none']",
"physical_beats": ["[beat one]", "[beat two]"]
},
"dialogue": [
{
"speaker": "[name/description]",
"delivery": "[tone, pace, pauses]",
"lines": "[exact dialogue text, with [pause] markers as needed]"
}
],
"audio": {
"ambience": "[environmental/room tone, or 'none']",
"sound_effects": "[specific diegetic effects and timing, or 'none']",
"music": "[style/mood/instrumentation, or 'none']",
"constraint": "The only audio in this video is the dialogue above exactly as written, plus the ambience/sound_effects/music explicitly listed above.",
"excluded": ["narration", "voiceover reading this description", "any audio not listed above"]
},
"exclusions": ["[unwanted visual or audio element]", "[unwanted visual or audio element]"]
}
PLAIN TEXT PROMPT TEMPLATE
Use for anything beyond a static one-shot talking-head: camera movement, style/era anchoring, multiple speakers, sound design decisions, or a sequence of beats/cuts. Keep every heading even when a category is unused — write "none" rather than omitting it, so the model treats it as a decision, not a gap. For a single continuous shot, skip the Timeline section and describe action directly under Scene. For multiple beats or cuts, use the Timeline section and leave Scene as a one-line overview.
Visual direction only, no narration read aloud.
Look: [medium/format, film stock or grade word, era, quality of light]. Presentation: [how/where this is being viewed or framed, e.g. projected in a theater, straight-to-camera, security footage, or "direct" if none].
Camera: [shot type/framing]. Movement: [static / push-in / handheld / pan / etc]. Movement change: [if the camera's movement shifts partway through, describe when and how, or "none"].
Scene: [one-line overview of who is in frame, the setting, and what is happening overall].
Timeline (only for multi-beat or multi-shot prompts — omit this section entirely for a single continuous shot):
[Beat 1, 0–Xs]: Camera: [shot/movement for this beat]. Action: [what happens]. Dialogue: [line spoken in this beat, or "none"].
[Beat 2, X–Ys]: Camera: [shot/movement for this beat]. Action: [what happens]. Dialogue: [line spoken in this beat, or "none"].
[Continue one beat per line, each with its own timecode, camera, action, and dialogue-or-none.]
Performance: Pacing: [slow/fast, deliberate/urgent]. Breathing: [timing/rhythm/breath notes, or "none"]. Physical beats: [ordered list of gestures/expression changes, or reference the Timeline above if already covered there].
Dialogue (list one entry per speaker; for a single speaker just use one):
Speaker: [name/description]. Delivery: [tone, pace, pauses]. Lines: "[exact dialogue text, with [pause] markers as needed]"
Speaker: [name/description]. Delivery: [tone, pace, pauses]. Lines: "[exact dialogue text]"
Audio — ambience: [environmental/room tone description, or "none"].
Audio — sound effects: [specific diegetic effects and when they happen, or "none"].
Audio — music: [style/mood/instrumentation if present, or "none"].
Audio constraint: The only audio in this video is the dialogue above exactly as written, plus the ambience/sound effects/music explicitly listed above. Do not generate narration, a voiceover reading this description, or any audio not listed here.
Exclusions: [explicit list of anything unwanted, visual or audio — e.g. no text on screen, no modern objects, no extra people, no score, no narration].
My favorite part of film editing is when you have a shot that doesn't work in the scene it was written for.
But you love that shot and you don't wanna lose it to the floor.
So you find someplace totally unexpected to use that shot, and like magic, it brings a whole new meaning to the scene.
Voice acting is what struggles the most in a gen ai performance right now.
You can pull really strong physical face performances out of characters with the right text.
Body language is close but still feels a little stiff. Meaning the best performances seem to be coming from close ups.
If you use caricatures it tends to work better because its mimickery at its finest. Less subtext and context needed when the character is familiar to models and the voice becomes more well scoped.
But I think the best voice acting in gen ai still comes from a human raw or with a voice change.
Voice is the key factor for an emotional performance, because that's where an actor completely reveals their truth. It's what we respond to on a primal level.
AI is only approximating the emotion. It does not draw on lived memories, where each memory contains complex sub-emotions, which can all be heard in a great actors performance.
At this point endlessly prompting to achieve results feels less value than hiring a voice actor and changing their voice and using it as a reference.
I hope more models begin to work with voice as a starting point. Because if analyzed correctly much of the desired physical performance can be derived from voice.
I'm impressed
all I asked our agent to do was:
"make a 20-second-long dialogue scene back and forth between two characters. Use the following prompt structure and fill in all of the required information. The setting should be a noisy street corner in New York City during the day and the two characters talking are a man in a window yelling down and a garbage man loading garbage into a truck."
Then I gave it prompt format - no creative details and let the agent come up with everything else.
What gets me is that the argument is actually coherent, that's the writer brain in our agent going to work - it understands conflict.
Don't mind the guy in the background endlessly throwing trash into the side of the truck - we're working on that 😂
🤖 Made with AI
I finally feel like I'm directing.
This short was 7 prompts and 2 reference images.
Here's the method...
Script is king. Write like a filmmaker. Have a story, name the emotion, time the beats, describe exactly what's in each shot and how you see it.
Seedance 2.5 is unmatched - this is the goto video model. All character design done in Midjourney.
Think in 30-second chunks. I wrote the script, had our agent time it out, then split it at safe cut points. No chunk ends mid-line or mid-action.
Then I designed every shot in the shot list for each chunk. I wrote far more subtext under the dialogue than I would in a normal script. The model got the character emotions right much faster.
Every chunk used the same prompt format from my other seedance vid posts.
References for each chunk: the previous 30-second video plus the two character images. The video carries audio, tone, and lighting. The character refs keep the looks locked.
Then into Resolve for editing. I had all the coverage I needed and the scene was 75% there. Adjusted some timing, reordered a few shots, cleaned up audio.
Three lines and one reaction shot needed a reshoot. Back into Preview, ref a single solid video plus the two character images, new prompt with just those pieces (about 15 seconds), cut it in.
Two drawbacks: cost and speed. Each gen takes a while and you need the previous chunk before you can start the next. And Seedance is not cheap. Against a traditional shoot though, still absurdly cheap and fast.
Have you tried something like this? What would you change?
😂
You’re unable to view this Post because this account owner limits who can view their Posts. Learn more
Short 5 of 100 - The Apology Machine
This final version is a mashup up of three different 30 sec generations from seedance 2.5
Each video was the same 30 second script using the same refs, but I changed the prompt slightly each time.
Heres the trick: use the same prompt, but change the description and length of specific shots that are giving you trouble.
By doing this you get the same general output, the same overall flow, BUT you can add emphasis to specific shots to then cut out and stitch together.
Can you spot which shots are from different vids?