Samurayich
374 posts
A 22 YEAR OLD COLLEGE DROPOUT DROPPED ONE $500 USED RTX 3090 INTO HIS BEDROOM PC AND NOW PULLS $6,200 A MONTH SELLING AD CREATIVES TO DROPSHIPPERS
leo is 22, lives in his parents place in tampa, was paying $400 a month on midjourney pro, runway and chatgpt to make tiktok ad creatives for 4 shopify clients
he grabbed a used 3090 off facebook marketplace for $500, slid it into his ryzen rig next to the old 5080
pause at 0:08 to see the thumbs up over the 24GB founders edition card, that is the moment the cloud bill goes to zero
runs flux dev for stills and wan 2.1 video locally, spits out 180 ad variations a day, charges $1,550 a month per client, four clients on retainer
$6,200 in, $9 in florida power out, the card paid itself back in 3 days, no midjourney queues, nothing leaves the bedroom
the window is open, follow and bookmark before it closes
starmex@starmexxx
English
New tongue-physics test for the best AI video models!
Seedance 2.0, Kling 3.0 Pro, Gemini Omni Flash, and Grok Imagine 1.5 were given a task that sounds ridiculously simple for a human:
Rotate the tongue in a continuous circle for 10 seconds. That’s it.
Each model received exactly four attempts.
Yet this tiny movement turned into an unexpectedly brutal test of instruction following, anatomy, motion consistency, and basic physical control.
And the result will probably surprise a lot of AI video creators. Spoiler: Seedance 2.0 completely failed to live up to its reputation this time.
- Kling 3.0 Pro:
The old man somehow managed to detect a policy violation in one attempt, although the other three generations went through without any problems.
Unfortunately, the result was still the worst of the four.
Kling failed the main instruction every single time: the tongue never actually rotated in a circle.
It moved sideways, twisted, stretched, and performed all kinds of strange anatomical experiments, but it never completed the requested motion.
The image itself looked reasonably realistic, but visual quality was almost irrelevant here. The movement was unnatural, inconsistent, and completely missed the task.
Everything that could go wrong did.
- Grok Imagine 1.5:
This was an unexpectedly good result.
The image still has Grok’s usual slightly cartoonish appearance, but its instruction following was surprisingly strong.
The tongue actually rotated.
Not perfectly, not with completely realistic anatomy, and not always through a clean full circle, but Grok understood what it was supposed to do and made a genuine attempt to execute it.
Considering its performance in several previous tests, this was a very pleasant surprise.
- Gemini Omni Flash:
The only model that handled the task convincingly.
Omni understood the instruction, produced a realistic-looking image, and generated an actual circular tongue movement.
It still isn’t perfect, and the motion does not yet look 100% like something a real person would perform. But compared with the other models, this was easily the cleanest, most controlled, and most believable result.
A genuinely strong performance.
Google finally found a very specific battlefield where Omni Flash can flex.
- Seedance 2.0:
This was the biggest disappointment.
The result was better than Kling’s, but Seedance still failed to perform the requested circular movement.
Seedance remains excellent at large action scenes, complex choreography, and cinematic motion. But in these simple human-body tests, including running and now tongue control, it keeps exposing unexpected weaknesses.
The movement looked more polished than Kling’s, but the core instruction was still not followed.
And for the first time in one of my tests, Seedance finishes below Grok.
- My ranking:
1. Gemini Omni Flash. The best instruction following, the cleanest circular motion, and the most realistic overall result.
2. Grok Imagine 1.5. Not perfect, but it understood the assignment and produced a surprisingly decent attempt.
3. Seedance 2.0. More polished than Kling, but clearly behind Grok and Omni in actual task execution.
4. Kling 3.0 Pro. Censorship, broken anatomy, and almost every possible tongue movement except the one I requested.
What do you think of this test? Do you agree with my ranking?
#AIVideo
Alpha Mom@YourAlphaMom
Another body-physics test for the best AI video models. This time, Elon’s pride and joy Grok Imagine 1.5 faces the old but stubborn Kling 3.0 Pro, China’s flagship Seedance 2.0, and America’s shiny newcomer Gemini Omni Flash, which developers called “revolutionary,” although I’m still waiting to see the revolution. The new battlefield: jump-rope exercises filmed from the front. There is a lot to analyze here. But let’s be honest. We all know exactly where everyone will be looking. Or rather, at which two points. The result was surprisingly close. - Kling 3.0 Pro: In this test, Kling was the only model that detected adult content in a completely harmless fitness scene and repeatedly refused to generate it. Which is especially funny considering how often Kling is used to bypass censorship with celebrity faces, yet apparently a woman jumping rope is where it draws the moral line. After several attempts, it finally produced a usable result. The lighting and overall realism are strong, as usual, but the actual body physics are the weakest here. It also missed the visual instructions from the prompt, giving me an older-looking model with less flattering proportions instead of the athletic, curvy character I requested. Not terrible, but definitely not Kling’s finest performance. - Grok Imagine 1.5: Apparently, Elon’s creation performs better from the front than from behind. Grok understood the task immediately, generated quickly, and didn’t complain about censorship. The image still has that unmistakably cartoonish Grok look, but the jumping motion is surprisingly fun and the body physics are actually decent. It feels more like a cutscene from a video game than real footage, but this is still one of Grok’s better results in my recent tests. - Gemini Omni Flash: Once again, Omni produced a beautiful and polished image. Google’s usual slow-motion, low-FPS effect is normally one of its biggest weaknesses, but in this specific test it actually worked in its favor. The slower movement makes the secondary motion much easier to see, and the body physics look smooth, convincing, and surprisingly natural. I liked this result a lot. - Seedance 2.0: Seedance performed much better than in the previous body-physics test. The footage is dynamic, vibrant, realistic, and visually the strongest of the four. The physics still don’t look completely natural to me, but the overall result is so convincing that it deserves first place, or at least a shared victory with Omni Flash. Omni may have slightly better body physics, while Seedance wins in movement, realism, energy, and overall image quality. - My ranking: 1. Seedance 2.0 and Gemini Omni Flash. A shared first place. Omni wins on body physics, while Seedance wins on overall realism, visual quality, and dynamic motion. If I absolutely had to choose one, Seedance would take it by a very small margin. 2. Grok Imagine 1.5. Still behind the two flagships, but noticeably better than in the previous tests. 3. Kling 3.0 Pro. Too much censorship, weaker instruction following, and the least convincing physics. The realistic lighting and natural-looking footage save it from being a complete failure. What’s your ranking? And if anyone wants the prompt, ask in the comments and I’ll share it. #AIVideo
English
📜The difference between Seedance 2.0 and Kling 3.0 is the difference between a thumbnail business and a watch-time business.
Pause 0:16
Same prompt. Two video models. One scene of a man in a penthous at nigth looking down at his city.
The result split into two completely different movies.
Seedance 2.0:
Establishing shot is the highest ceiling I've seen this year. The skyline at night looks like the opening frame of a Christopher Nolan film — depth, density, real city lights reflected on river water. When the camera moves into the apartment, the lighting holds. Furniture catches the warm lamp glow correctly. The actor sits with weight, not pose. Frame by frame this is photography, not generation.
Kling 3.0:
Where Kling pulls ahead is the moment of transition. The elevator shot — doors opening, the man stepping forward — has body mechanics no still frame can fake. Shoulders shift before the leg. Hand reacts to balance. The camera doesn't tell you he just arrived, the physics does. Some background blur, some weird ceiling artifacts, but the human inside the frame moves like a human.
The honest take:
Seedance wins the still you'd put on a billboard.
Kling wins the second you press play.
If your channel is built on thumbnails — Seedance.
If your channel is built on watch time — Kling.
The "best model" question is dead. The "best model for this shot" question just opened.
Which one would you have published? 👇
Samurayich@Rulia2505
New text-to-video battle for yhe best AI generators: Recently released Grok Imagine 1.5, the king Seedance 2.0, and the old veteran King 3.0 This time I tested hiw each model handles an apocalypse scene - lighting storms, collapsing cities, first-person driving though chaos. Real physics, real weight, real atmosphere. And the result was much less obvious than I expected. Believe it or not — this time I can't call Seedance a clear winner. Each model got several attempts. — Seedance 2.0: Cinematic as always. Lightning, broken spires, glowing storm — straight out of a movie poster. The composition is stunning, but it feels like a still frame in motion. Beautiful, but emotionally distant. You watch it like art, not like a scene you're inside. — Kling 3.0: The dark horse of this test. Best rain physics on the windshield, best sense of speed, best "I'm actually driving through hell" energy. The dashboard reflection alone sells the entire shot. Some weird hand artifacts on the wheel, but the atmosphere is undefeated. — Grok Imagine 1.5: The biggest surprise. Less dramatic, less cinematic — but somehow the most real. Smoke, fire, debris all sit naturally in the frame. It doesn't try to impress. It just looks like a Tuesday in a war zone. That restraint is what makes it hit harder. — My ranking: Kling 3.0 — best physics, best immersion, best "you are there" feel Grok Imagine 1.5 — surprisingly grounded, scariest in the quiet way Seedance 2.0 — gorgeous, but felt like a render, not a moment Agree with my ranking? 👇
English
@JonBuildsHQ No cheating. The reply game is a skill - you just automated the boring half. Ship it 👇
English
I created this tool to grow faster on X.
Is this cheating? 😂
Should I release this to the public?
English
@starmexxx $400 of ex-mining GTX 1080s replacing a $400/mo cloud sub is the cleanest arbitrage of the decade 📉 ETH leaving PoW accidentally funded the local AI revolution
English
STACKING 4 OLD NVIDIA GTX 1080s GIVES YOU LOCAL AI THAT BEATS $400/MO IN CLOUD SUBSCRIPTIONS
people are pulling 4 used gtx 1080s off ebay for around $400 total and stacking them in one tower to build a local ai rig. that pools 32gb of vram across cards that originally retailed at $700 each
old gaming cards became cheap when ethereum killed proof of work in 2022. miners dumped millions of them and developers picked them up to run open weight models locally
ollama installs in one shell command and pulls a model in another. point claude code at localhost through one environment variable and the cli works identically without sending data to anthropic
llama 3.2, mistral 7b, qwen 2.5 and phi 3 all run comfortably on a 1080 stack at 40 to 60 tokens per second. fast enough for real conversations, automation pipelines and document analysis
a heavy ai user pays $200 chatgpt pro, $200 claude code and $20 cursor every month. that's $5,040 a year and a $400 rig pays itself off before month two
bookmark this and read the article below
Noisy@noisyb0y1
English
@andreysuperior 50 faceless accounts > one store > digital products that deliver themselves is the cleanest passive funnel I've seen described in month
English
A girl on a balcony made $1,700 today from silly little TikToks she never appeared in.
She said it doesn't feel real. Then showed the dashboard.
50 faceless accounts. No face. No voice. Claude wrote every script. A robot voice read them. Buffer scheduled them.
Every video sends viewers to one store. Digital products. They deliver themselves.
5 hours of work a week. $57 a month in tools.
The dashboard says $1,700 today.
She's standing over the city like she still can't believe it.
Superior@andreysuperior
English
📜Two video models. One impossible scene. A guy riding a glowing manta-ray vessel - through a foggy forest into a roaring waterfall canyon.
Ran the same prompt through Seedance 2.0 and King 3.0
Pause 0:21
The result wasn't what I expected.
Seedance 2.0:
Visually unreal. The starlight texture on the wings, the warm thruster glow underneath, the way the forest mist sits around the vessel — every frame looks like a Marvel concept artist had a perfect day. The downside? It feels too composed. The rider doesn't move. The world doesn't react. You're looking at a painting that happens to be in motion.
Kling 3.0:
Less polished, more alive. The waterfall scene is the moment it pulls ahead — actual water spray reacting to the craft, real sense of weight cutting through air, body shifting with the turn. Some frames blur into abstract paint, but when it works, it works hard. This isn't a poster. It's a clip from a movie I'd actually watch.
Verdict:
Seedance wins still frames. Kling wins motion.
If you're making thumbnails — Seedance.
If you're making scenes — Kling.
The era of "one model wins everything" is over. Pick the tool for the shot, not the brand.
What's your call on this one? 👇
Samurayich@Rulia2505
New text-to-video battle for yhe best AI generators: Recently released Grok Imagine 1.5, the king Seedance 2.0, and the old veteran King 3.0 This time I tested hiw each model handles an apocalypse scene - lighting storms, collapsing cities, first-person driving though chaos. Real physics, real weight, real atmosphere. And the result was much less obvious than I expected. Believe it or not — this time I can't call Seedance a clear winner. Each model got several attempts. — Seedance 2.0: Cinematic as always. Lightning, broken spires, glowing storm — straight out of a movie poster. The composition is stunning, but it feels like a still frame in motion. Beautiful, but emotionally distant. You watch it like art, not like a scene you're inside. — Kling 3.0: The dark horse of this test. Best rain physics on the windshield, best sense of speed, best "I'm actually driving through hell" energy. The dashboard reflection alone sells the entire shot. Some weird hand artifacts on the wheel, but the atmosphere is undefeated. — Grok Imagine 1.5: The biggest surprise. Less dramatic, less cinematic — but somehow the most real. Smoke, fire, debris all sit naturally in the frame. It doesn't try to impress. It just looks like a Tuesday in a war zone. That restraint is what makes it hit harder. — My ranking: Kling 3.0 — best physics, best immersion, best "you are there" feel Grok Imagine 1.5 — surprisingly grounded, scariest in the quiet way Seedance 2.0 — gorgeous, but felt like a render, not a moment Agree with my ranking? 👇
English
@gippp69 Kimi K2.6 executes, Opus 4.8 audits - that's the actual multi-agent architecture nobody's talking about yet 🔥 cheap muscle + expensive judgment is the winning stack
English
A CHINESE DEVELOPER JUST USED KIMI K2.6 TO RUN 300 AI AGENTS, 4,000 RESEARCH STEPS AND 5 LIVE DATA FEEDS WITHOUT LETTING THE SYSTEM SHIP BAD NUMBERS.
most agent swarms are fast, but they are also dirty. this setup runs differently. Kimi K2.6 executes the swarm, Claude Opus 4.8 checks the work, rejects broken outputs, and sends failed tasks back into the loop.
he tested it on 100 EV companies. the goal was not a random AI summary, but a research-grade report with a comparison matrix, margins, revenue numbers, citations, and every figure traced back to a live source.
the first pass came back in minutes, but it was not clean. Opus rejected 12 companies. some had revenue numbers that did not match the feed, two had dead citations, and one had an empty margin field.
those 12 were pushed back into the queue with the exact reason attached. second pass had 3 failures. third pass had zero. the loop stopped only when nothing was left to reject.
that is why Kimi matters here. raw swarms just scale the error count. Kimi K2.6 gives the system parallel execution, while Opus turns it into something closer to an analyst team with a built-in auditor.
300 agents are cool. 300 agents that check themselves are the real unlock.
0xRicker@0xRicker
English
@qusaxz Step 1 is the whole game. Finding the pain > knowing the code. The "no-code" era is really the "good taste" era 🔥
English
He sells custom apps to businesses for thousands of euros.
He’s never written a line of code.
Here’s his 30-second method:
Step 1: Ask Claude what problems exist in one specific niche.
Restaurants. Construction companies. Doesn’t matter which.
The goal is finding pain points AI can actually solve - something that saves a real business real time or real money.
Step 2: Ask Claude to write one detailed prompt.
A prompt big enough to describe the entire app - the problem, the solution, the design, the flow.
He copies it.
Step 3: Feed that prompt into an AI design tool.
Out comes a clean, polished interface. The kind that looks like a funded startup built it.
Here’s why businesses pay for this.
They’re not paying for code. Code is commoditized now.
They’re paying because someone walked in with a working solution to their specific problem, built in an afternoon instead of a six-month dev cycle.
The skill that matters isn’t programming anymore.
It’s asking AI the right question about the right pain point.
He just turned that into a several-thousand-euro service.
English
@starmexxx @Creatify_AI "Convert was never the bottleneck. Conversion was." - that opening line does more work than most landing pages 🔥
English
CONTENT WAS NEVER THE BOTTLENECK FOR D2C BRANDS. CONVERSION WAS. THIS IS THE FIRST AI BUILT FOR THAT GAP
@Creatify_AI connects to your meta, tiktok, google, shopify and ga4 accounts in two minutes and studies what's already winning in your niche
the agent flags wasted spend, generates new creatives based on what actually converts, and launches them across every platform you run ads on
emails, reports and the tedious work between strategy and launch are handled inside the same chat interface
two minutes to connect, no credit card to start. $39 a month after the trial
Creatify AI@Creatify_AI
AI Media Buyer is live in Creatify. Most ad accounts run on gut. Yours should run on data. • Connect - Meta, Google, TikTok, and the accounts you already run • Analyze - Winners, losers, and the spend quietly leaking • Act - The agent doesn't just flag the move. It makes it. • Scale - New creative built from what's already winning in your niche Try it now in Creatify.
English
@cryptansky The "no face, no voice, no girl" line goes hard. Whole AI era in one breath 👀
English
Two girls just played on your screen, side by side.
One was real. One was a deepfake.
Most people pick wrong.
Here is the part that should actually bother you.
It does not matter which one you picked.
A 23 year old in Tallinn runs one of these.
$41,000 in its first 31 days.
2,180 men pay to talk to a girl named Sofia.
Sofia is 24. A photography student from Lisbon.
Sofia is five text files in a folder on a laptop.
No face. no voice. no girl.
Claude writes every message.
Flux generates every selfie.
ElevenLabs sends the 2 a.m. voice notes.
The whole company costs under $200 a month.
The man who sends her $2,300 a month is not stupid.
He is not even fooled, exactly.
He decided that something that feels like being chosen
was close enough.
You just spent ten seconds trying to spot the fake face.
The real trick was that there was never a face to begin with.
The next Sofia is a weekend away.
Full breakdown below.
CRYPTANSKY@cryptansky
English
Three video models. One impossible scene. A father shielding his kid from a giant troll in a forest, then King Kong hold-up shot. Pure fantasy chaos.
Pause 0:15
Ran the same prompt through Kling 2.6, Hailuo 2.3 and Seedance Pro.
Honestly thought I knew who'd win this one. I didn't.
Kling 2.6:
Composition is solid but it tried to do too much at once. Faces blur out in the action shot, the troll's anatomy gets weird in close-up, and there's that glassy "I'm a rendered demo" sheen on everything. Strong cinematography, weak humanity.
Hailuo 2.3:
The dark horse. Lighting feels like an actual forest in actual fog. The way the father pulls his kid behind the rock has real protective body language — not posed, just reactive. The King Kong shot is also the most physical of the three. You can feel the weight. Some background flicker but the emotion lands.
Seedance Pro:
Cleanest faces, cleanest detail, but emotionally the flattest. Kid looks like he's posing for a poster, not running from a monster. The Kong scene is gorgeous but reads like fantasy art, not a frame from a film. Beautiful — just not scary.
My ranking:
Hailuo 2.3 — emotion, weight, fear. Closest to a real scene.
Seedance Pro — visually richest, narratively weakest.
Kling 2.6 — strong frames, but the action breaks it.
Hailuo quietly cooking while everyone's still arguing about Seedance vs Kling 👀
What's your top 3?
English
@gippp69 $10k ones to never pay OpenAI or Anthropic again is the new mortgage 🏠 521 GB local rack going to print money quietly for years
English
HE STACKED 4 MS-S1 MAX-STYLE BOXES IN A SERVER RACK. 4 × $2,599 TO STOP PAYING $10-$20 EVERY DAY FOR AI AGENTS.
each box is built around Ryzen AI Max+ 395, 128GB unified memory, and up to 96GB shared GPU memory. four of them turns a small rack into 512GB of local AI infrastructure.
the stack is simple. Hermes Agent runs the workflows, llama.cpp serves the model locally, and the system points to localhost instead of renting OpenAI or Claude tokens all day.
the economics are the real point. $10-$20/day looks small until agents are running research, drafts, summaries, client docs, tool calls, and sub-agents every single day. that becomes $300-$600/month for repeat work.
one MS-S1 MAX can run a 120B local model around 56 tokens/sec at roughly 110W under load. with four boxes, you can split writing, research, automation, dashboards, and backup jobs across the rack.
for small agencies, content teams, research shops, and solo founders, this changes the math. the agent can work 24/7 without charging you every time it thinks, reads, writes, or touches a file.
local AI is not a chatbot on a desk anymore. it is owned infrastructure in a rack.
NO1ennn@N01ennn
English
@qusaxz "Build, runs ads, sell leads" is the leanest agency model I've seen all year. 5 clients > 50 clients hits different 🔥
English
2 guys from Switzerland are closing $2,000 website deals in Phuket right now
Not remote. Not “digital nomad lifestyle content”
Actually working
They use Bolt dot new to build the sites fast
They run Meta ads on top to generate leads for local businesses
Then they sell those leads back to the same businesses
Its a 3-part machine - build, run ads, sell leads
You dont need 50 clients
You need 5
5 businesses paying you monthly to keep the leads coming
Real estate is their main target
One sold house covers your invoice for 6 months
They started 6 years ago with dropshipping
Small websites, trial and error
3 years ago they registered a company
Now theyre sitting in Thailand explaining the model to strangers at a cafe
They work only in Switzerland right now
But the model has no borders
You build the site once
The ads run without you
The leads close while youre in Phuket
Thats not freelancing
Thats a system
English
@RoundtableSpace Diffusion for text was a meme 12 month agom Now it's breaking inference records weekly. Wild speedrun 🔥
English
Diffusion Gemma just hit 2000+ tokens per second locally - 1.8x faster than before, running on 18GB RAM.
Diffusion models for language are moving faster than anyone expected.
LOCAL INFERENCE SPEED RECORDS ARE BEING BROKEN WEEKLY NOW.
English
New text-to-video battle for yhe best AI generators:
Recently released Grok Imagine 1.5, the king Seedance 2.0, and the old veteran King 3.0
This time I tested hiw each model handles an apocalypse scene - lighting storms, collapsing cities, first-person driving though chaos. Real physics, real weight, real atmosphere.
And the result was much less obvious than I expected.
Believe it or not — this time I can't call Seedance a clear winner.
Each model got several attempts.
— Seedance 2.0:
Cinematic as always. Lightning, broken spires, glowing storm — straight out of a movie poster. The composition is stunning, but it feels like a still frame in motion. Beautiful, but emotionally distant. You watch it like art, not like a scene you're inside.
— Kling 3.0:
The dark horse of this test. Best rain physics on the windshield, best sense of speed, best "I'm actually driving through hell" energy. The dashboard reflection alone sells the entire shot. Some weird hand artifacts on the wheel, but the atmosphere is undefeated.
— Grok Imagine 1.5:
The biggest surprise. Less dramatic, less cinematic — but somehow the most real. Smoke, fire, debris all sit naturally in the frame. It doesn't try to impress. It just looks like a Tuesday in a war zone. That restraint is what makes it hit harder.
— My ranking:
Kling 3.0 — best physics, best immersion, best "you are there" feel
Grok Imagine 1.5 — surprisingly grounded, scariest in the quiet way
Seedance 2.0 — gorgeous, but felt like a render, not a moment
Agree with my ranking? 👇
English
@Bober_smart "Owner thinks he spent a week on it" - the perception gap is where the margin lives 🤣
English
A 24-year-old guy from Britain turned himself into a girl and works as a psychologist for men, charging $500 an hour
His only expense is $20 for Claude
He put up some ads, and men lined up in a queue. During the consultations, they pour their hearts out and share their problems
Claude alters his appearance and voice in real-time
He conducts 15 consultations a week = $7,500 a week
And the funniest part is, they all know it’s AI
Bober_smart@Bober_smart
A 24-year-old guy from Texas turned himself into Elon Musk and made $4,356 in 4 hours It all started as a joke, he launched a TikTok stream, but the result was unpredictable The costs amounted to $20 for Claude Claude changed his appearance in real time The audience loved the look so much that they started donating and sending gifts At the end of the stream, his balance was $4,356 Peak viewership reached 11,430 The stream gathered 186,580 likes He gained 65,000 new followers And the funniest thing is, everyone knew it was AI
English
@YourAlphaMom Body physics is the new benchmark - anyone can fake a still frame, nobody fakes a sprint 🏃♂️curious which one cracked the cloth sim
English
New body-physics test for the best AI video tools:
The newly released Grok Imagine 1.5, the king Seedance 2.0, Google’s “revolutionary super-duper” Gemini Omni Flash, and the old-timer Kling 3.0 Pro.
This time I tested how each model handles realistic running motion, body movement, fabric physics, and natural secondary motion.
And the result was much less obvious than I expected.
Believe it or not, I can’t call Seedance the clear winner this time.
Each model got several attempts.
- Grok Imagine 1.5:
The new version is finally available on the official Grok website, so I could properly test it.
Honestly, I didn’t notice a massive improvement over the previous version. Still, the result was acceptable.
It produced the most cartoonish image, and the woman runs as if she’s wearing heels, but the body physics were decent enough. It also understood the instructions quickly and followed them correctly.
- Kling 3.0 Pro:
The old man decided to test my patience.
It repeatedly blocked a simple running scene as adult content, then misunderstood the instructions several times.
The successful result has the most realistic lighting and frame rate, but the actual body physics look strange. It almost feels like loose foam padding is bouncing inside the leggings.
There are also several visible artifacts and unnatural movements.
- Gemini Omni Flash:
As usual, it gave me that strange slow-motion, low-FPS look that Google models seem to love.
But it didn’t censor anything, understood the instructions immediately, and produced a beautiful, realistic result.
Surprisingly, this is the output I liked the most in this test.
- Seedance 2.0:
Seedance also blocked a couple of generations, just like Kling, but eventually produced a strong result.
It delivered the most beautiful and visually appealing footage, but I honestly expected better physics.
The video looks great, yet I can’t confidently call it the most physically accurate result.
- My ranking:
1. Gemini Omni Flash — not perfect, but the best overall result for me
2. Seedance 2.0 — visually stunning, but I preferred Omni’s physics
3. Grok Imagine 1.5 — cartoonish look and strange running, but still acceptable
4. Kling 3.0 Pro — the longest wait, the strangest physics, and the most inconsistent result
Do you agree with my ranking?
#AIVideo
Alpha Mom@YourAlphaMom
New action test: Seedance 2.0 vs Gemini Omni Flash vs Kling 3.0 Pro vs Google Veo 3.1. This time, a Lara Croft-style heroine had to escape armed enemies, perform multiple stunts, and shoot back during one complex action sequence. Each model got 4–6 attempts. I brought Veo 3.1 back instead of Grok Imagine. The reason is simple: Veo handled this scene better, while new Grok 1.5 still isn’t available on the official Grok website, so I left it out. - Google Veo 3.1: Predictably finished last, but I wouldn’t call the result terrible. It clearly looks weaker and more outdated than the others, but for such a difficult scene, the result was still acceptable. - Gemini Omni Flash: Not perfect, but still pretty good. The stunts, physics, and overall execution worked well. My main issue is Google’s familiar low-FPS look. The footage constantly feels slightly slowed down, which makes it less cinematic than both Seedance and Kling. Still, I liked the result, and it was strong enough to take second place. - Kling 3.0 Pro: Better than I expected. Its biggest problem is character consistency. The heroine and other characters become distorted during fast movement, with plenty of unnatural poses and animation errors when you look closely. The overall result is still acceptable. Worse than Gemini Omni Flash, but definitely better than Veo 3.1. - Seedance 2.0 This was Seedance’s territory. Complex action remains its greatest strength, and it started producing excellent results within the first few attempts. The physics, movement, instruction following, and overall intensity were all impressive. This might be one of the best results I’ve ever received from Seedance. This test once again shows that Seedance remains the king of difficult action scenes. - My ranking: 1. Seedance 2.0 — clearly on another level 2. Gemini Omni Flash — strong, but held back by the slow-motion feel 3. Kling 3.0 Pro — better than expected 4. Google Veo 3.1 — predictably failed again What do you think of the new test? #AIVideo
English