The last ninety days collapsed generative media into a single gesture: you hand over one photograph and describe a physical event. Here are 100 effects that actually spread — sorted into seven techniques, each with the real artifact, a colour-coded breakdown of the prompt, and the failure modes that kill it. Not a gallery to scroll. A lab to work in.
Every effect on the board is one of seven things. Choose the kind of control you want and the board filters to it. If you are not sure where to start, take the tour — it spends its time on the two things people get wrong.
Two axes, not one list. The seven techniques above are what an effect is — you browse them. These seven controls are what you set, on every single item, whatever the technique. Learn the controls once and you can rebuild ninety per cent of the board below.
Seven techniques — what kind of effect is it
One per item. Drives the skill map and the gallery filter. These are content categories, and they are what you browse when you do not yet know what you want to make.
Seven controls — what you set, every time
Universal. Applies to every item regardless of technique. These drive the prompt anatomy, Change One Thing, and the Prompt Builder. They are not places to browse.
One clear, well-lit photo. Face forward, no filters, no sunglasses, no group shot. If the effect involves footwork or a full-body routine, use a full-length shot — a waist-up photo gives the model nothing to stand on.
Photo → cinematic clip with native audio: Seedance 2.0/2.5. Dance and gesture routines: motion control (Kling, Seedance). Camera-motion effects: Higgsfield presets. Stills and edits: GPT Image 2.5 or Nano Banana Pro.
9:16 vertical is the default and the algorithm's preference. Go 16:9 only when the format demands it — bullet time, a windshield two-shot, a fisheye pool selfie. State the aspect ratio in the prompt or you'll get it wrong.
"Cinematic" produces nothing. Name the camera, lens, light and the physical beat: handheld phone footage, 30–40° orbit, hard side light, one unbroken take. The physics is the effect.
Modern models generate audio in the same pass — squelches, rotor wash, crowd noise, room tone. For voice models, write bracketed performance directions. For trend audio, the sound is the discovery mechanism; the caption names the artist, not the meme.
Most of these effects are defined by what you forbid: no cuts, one take, keep the real background, don't tidy the room, seal the mask shut. Continuity is the trick — an edit breaks the illusion instantly. It is a constraint, not a category: nothing is "a continuity video".
Face smear on motion transfer. Hands melting when arms cross the body. A subject that stays swollen instead of springing back. A hand that's merely large. A figurine that turns human. Build the check into your workflow.
One hundred effects. Click any card for the full lesson: the real artifact, the prompt taken apart control by control, Change One Thing, the failure modes and a quiz on how to fix them. Cards marked ▶ Real example show a genuine video of that effect — linked and attributed, never rehosted; the rest show a hand-drawn diagram of the mechanics.
If you would rather be told what to look at, in what order, with a reason each time — pick a path. Each one is a short sequence of real items from the board, chosen because together they teach something the individual entries do not.
The tools behind the board. Everything dated 3 July 2026 or later sits inside the research window; the two older entries are listed because the wave of the last 90 days runs on them.
Prompt galleries are the main way image techniques actually reach creators — and they have a rigid, recognisable structure. Learn to read it and you can extract the transferable part in about ten seconds, instead of saving a post you will never open again.
"20 Award Winning AI Image Prompts", delivered as an Instagram carousel and mirrored to a Substack. The Substack positions itself as a library of 10,000+ image and video prompts plus 30+ hours of training. The free post shows the intro; the prompt bodies sit behind the subscription.
Good for: seeing which prompt shapes are being taught, and how a gallery is packaged. Not for getting the prompts.
Publishes a short, dated list of what is trending that month with the full prompt text inline and no paywall. The August edition is the best source I found for genuinely copy-paste-ready prompts inside the window.
Good for: the actual text. Several entries on the board below are lifted from here.
The broadest catalogue — 100 prompts across ten categories, refreshed weekly. It is also the source that explicitly names the 1980s retro wave as the dominant trend cluster of late 2026, with the regional variants that made it last.
Good for: breadth and the dated trend call. It does not state aspect ratios or per-prompt models, so you supply those.
A large collection that is unusually strong on the unglamorous prompts — background replacement, object removal, natural skin retouching, and the "my face doesn't match the photo" fix. The troubleshooting append-ons are the most valuable part of the page.
Good for: corrective prompts. This is where you go when a generation is nearly right and you need the clause that fixes it.
The only source I found doing month-by-month creator mechanics with real failure modes. It also sells a song generator, and says so inline — which is more disclosure than most of this ecosystem manages.
Good for: method. The texts-to-song guide is the best-written recipe in the whole research set, and it works with any generator.
Not a prompt gallery — a preset gallery. One-click motion effects with "Top Choice" badges and no prompt to read at all: Earth Zoom Out, Bullet Time, Red Carpet, Disintegration, transformation effects. The prompt has been absorbed into a button.
Good for: speed, and for seeing where this is heading. Once a workflow becomes a button it stops being a trend and becomes a feature.
Everything below is the working method behind the board — how the artifacts were found and rejected, how the evidence tiers are derived, and where this page is weakest. It is kept in full because it is the most unusual part of the page, but it is deliberately out of the way of the learning. Switch back to Learner at the top right to hide it again.
No media is downloaded, copied or rehosted anywhere on this page. Only a video ID is stored; the thumbnail is loaded from YouTube's own image CDN and the link sends you to the platform to watch it. Embedding and thumbnail hotlinking are the mechanisms YouTube provides for exactly this purpose, and linking to a publicly available work is not reproduction. TikTok posts are linked, never embedded and never mirrored, because their thumbnail URLs are signed and expire — a cached copy would break or, worse, silently show the wrong frame. Where no match passed, the panel gives you TikTok, Instagram and YouTube search links instead, so there is always a route to the canonical original. If you own a clip here and want it removed, the link is right there — say the word and it goes.
On the previews. Where a real video of an effect exists, the card and detail panel show that video's own thumbnail, hotlinked from YouTube's CDN, with the video's real title. Where it does not, the panel falls back to a stylised, hand-drawn SVG diagram of the mechanics — labelled as such, because it is not AI output and not the original.
Eighty-six of the hundred entries carry a real example — eighty-five a YouTube video, one a canonical TikTok post. Each one was found by search and then read by a human to confirm it actually shows the effect — a stricter bar than keyword matching, which happily returns the underlying song, the original film scene, or an unrelated topic that happens to share words. Eighty-five candidates were read and rejected on that basis and are absent on purpose — forty-six across the first four search passes, thirty-five across the next three, four in the gap-filling pass — and two artifacts that had already shipped were withdrawn: one for showing the wrong thing, and one whose video no longer had a real thumbnail (it served a solid black frame, so the card rendered as a black box). A wrong or blank image under a right headline is worse than no image, because it borrows credibility it has not earned.
The fourteen without an artifact fall into two groups. Twelve are TikTok-native memes and dance trends: for these the named trend itself is the only thing that qualifies, and a search returns the underlying song, a generic motion-control tutorial, or nothing at all — TikTok's own search endpoint is not reachable from the tooling used to build this page, which is where those trends actually live. The other two are image styles (risograph print, disposable-camera snapshot) whose technique tutorials do exist but are Photoshop and Lightroom recipes rather than AI ones; since every artifact on this page is AI-branded, attaching a non-AI recipe would misrepresent the entry.
Of the eighty-five video artifacts, thirty-two were uploaded inside the research window and fifty-three before it — some as far back as 2023. They are still the effect, and a tutorial from 2025 is often a better teaching example than a clip from last week, but they are not evidence that the effect trended in the last ninety days. Rather than bury that, every artifact prints its upload month on the card and in the detail panel, so you can weigh it yourself. Where a canonical post exists it is linked beneath the video. A dedicated pass searched specifically for in-window replacements for the weakest of them and found almost none: for these effects the tutorials that exist are simply older than the window.
Eighty-four of the hundred entries rest on a single source. Forty-one of those trace to one roundup (Starrd's monthly viral-AI-video-trends posts) and eight to one prompt-gallery site. That is a real methodological problem, not a footnote: a trend roundup is a secondary source, so if it misdates or overstates something, this page inherits the error. So the evidence tier is derived from the sources, not hand-written — an entry is only labelled Primary source if at least one of its links is a primary document, a first-party product page, or independent editorial. Everything resting solely on roundups or affiliate listicles renders as Reported. Current split: 13 primary, 79 reported, 7 vendor-described, 1 outside-window. Where a primary source exists it is attached as a second link — ByteDance's own Seedance 2.0 model card sits behind the five entries whose central claim is that audio is generated in the same pass as the video, and it is what corrected this page's earlier misdating of that model. Treat view counts, revenue figures and "X million views" claims as reported, not verified.
Replication steps are reconstructed from published creator workflows where available, and inferred from the mechanics where not. Where a step is inferred, the panel says so. The prompt anatomy added in this version is authored by hand, one item at a time, from the creator's own prompt — it is a reading of that prompt, not a machine parse of it. That is deliberate: measured against the raw text, only 66 of 100 prompts name their subject explicitly, 26 name the shot, and 17 name the motion, so an automatic parser would leave most cards looking broken. Where an entry's prompt is a workflow, a meta-brief or a template rather than one codable instruction, the anatomy is withheld rather than invented — that applies to eight entries, and the lesson says so in place of the breakdown.
Several sources that surfaced during research were discarded rather than cited: SEO-filler "AI sound directories", paywalled newsletters offering only a teaser, and at least one viral-video roundup whose entries could not be traced to any real post — it listed generic titles with implausible view counts and attributed the whole 2026 wave to models that were not the actual drivers. A trend list that cannot be checked is worse than a shorter one that can.
Set the seven controls, or start from one of the 92 worked examples and edit it. The scaffold below assembles itself as you go — copy it into any engine.
For running this live in front of a room.
Scan to open the lab on a phone. The QR points at the live site, so it keeps working after this session.
Suggested live flow: open a random lesson with prompts hidden, show the artifact, ask the room to describe the seven controls out loud, then reveal the anatomy and compare. The Change One Thing dropdown is the best two-minute exercise — swap the lighting or the motion and ask what breaks.