We build #AI products where creativity and innovation meet: stock image search engine, #texttospeech, #faceswap, lipsync, algorithms to moderate #digitalcontent
Joined June 2016
- Tweets842
- Following1.3K
- Followers1.5K
- Likes1.6K
Free vs paid AI upscaler. Blind test.
SeedVR2 (free, open weights) vs Topaz Wonder 3 (paid).
264 image pairs, 13 pros: 54–60% ties.
Coin flip on most images.
One exception: faces. Topaz won naturalness 52% to 14%.
#SeedVR2 #TopazLabs
research.everypixel.com/seed…
2
1
87
Think you can spot an AI image?
In a 2025 study, 104 people sorted 50 images into real and AI.
They caught the AI ones 63.7% of the time.
For FLUX.1-dev images: 29%.
So why don't we trust AI images yet?
#FLUX #Midjourney
journal.everypixel.com/trust…
37
GPT Image 2 passed 33 of 34 real production briefs.
The one miss: a Tokyo crowd scene, 6.9/10.
The catch: ask for 9:16 or 1:1 in the prompt and you still get 16:9. All 34 outputs did.
Set the size in the API call.
#GPTImage2 #OpenAI
research.everypixel.com/gpt-…
2
3
112
Lip sync is the easy part. The rest of the face is where AI avatars break.
We blind-tested 4 talking avatar models. Speech sync tied in 22 of 36 judgments. On expression and motion, MiniMax H3 Turbo won 75%. LTX 2.3 won 11%.
The winner costs half as much as the standard H3.
Full blind test: journal.everypixel.com/best-…
2
311
Seedance 2.0 vs MiniMax H3. Same six-panel anime storyboard.
MiniMax shot the panels in order.
Seedance improvised. Its titan came back with two heads.
MiniMax also costs us a third as much.
Our pick for storyboards: MiniMax H3.
#MiniMax #Seedance
journal.everypixel.com/seeda…
1
1
107
Extra fingers won't save you anymore.
In a 2025 study, people spotting AI images vs real photos scored ~54%. Basically a coin flip.
What still works: #SynthID watermarks, #C2PA Content Credentials and tracing the original source. Your eyes? Not proof.
journal.everypixel.com/how-t…
One upscaler restores. The other rewrites.
SeedVR2 vs Magnific. 792 blind votes from 13 pros.
SeedVR2 wins a random pair with 78% probability.
Faces, detail recovery: SeedVR2 93% of pairs. Magnific: none.
And SeedVR2 is free.
#SeedVR2 #Magnific
research.everypixel.com/seed…
34
Seedance 2.5 refused 3 of our 15 briefs. Kling 4 shipped all 15.
Same prompts, 10s 720p clips with audio:
#Kling4: 15/15, $1.57 a clip
#Seedance 2.5: 12/15, $3.60 a clip
Less than half the price, zero refusals. Seedance still wins on title text.
journal.everypixel.com/kling…
3
123
Can AI replace human image reviewers?
We gave 3 AI judges 264 image pairs our reviewers had already scored.
On the easy pairs, where all 3 humans agreed, the best AI panel still got 18% wrong.
Cost: $12. Verdict: humans stay.
#Gemini #GPT5
journal.everypixel.com/can-a…
50
Qwen Image 2.0 Pro: worth the upgrade?
48 scenes, 13 pros, blind.
Vs Qwen Image 2: Pro wins 63.0% on attractiveness, loses 21.5%.
Vs GPT Image 2: Pro wins 19.7% on believability, loses 50.6%.
Real upgrade. Not top tier yet.
#Qwen #GPTImage2
research.everypixel.com/qwen…
3
174
6 AI video models. One job: turn an evening shot into daylight.
The cheapest, at ~27 cents a run, moved the shadows right.
One costing 5x more moved the light and left every shadow in place.
Price told us almost nothing.
#FLUX #Gemini #Seedance
journal.everypixel.com/6-vid…
1
219
“Commercial use allowed.” Who pays if you get sued?
We checked 14 AI image offerings:
#GettyImages: up to $50,000 per image on standard purchases
#Midjourney: no general indemnification
#GPTImage2: covered on eligible API/Enterprise use, with exclusions
journal.everypixel.com/ai-ge…
1
126
OpenAI's new flagship vs the old one on realism? Coin flip.
GPT Image 2.5 Sunburst got 52.7% of 431 blind votes vs #GPTImage2.
vs #Qwen Image 2: a blowout, 69.0%.
Where Sunburst edges ahead: prompt adherence and aesthetics.
#GPTImage
research.everypixel.com/gpt-…
361
Tell people it's AI and they like it less.
2026 CHI study: sites with AI images won 60% of picks (153 of 255).
After disclosure: 133 originals, 125 AI.
Same pixels. Different verdict.
Our blind test, #GPTImage2 vs real photos: coin flip.
#Qwen #FLUX
journal.everypixel.com/ai-vs…
438
“Commercial use” doesn't mean you're covered.
We read 10 AI video licenses. Catches:
#MiniMax H3 open weights: not licensed in the US, EU, UK, South Korea
#LTX: free under $10M revenue, must disclose AI
#Runway standard plans: can train on your footage
journal.everypixel.com/ai-vi…
196
Seedream 5 Pro beats Qwen. But GPT Image 2 still beats Seedream on realism.
Blind test, 13 pros, 48 scenes:
vs #GPTImage2: 91 believability wins, 204 losses
vs #Qwen Image 2.0: 284 aesthetics wins, 88 losses
#Seedream
research.everypixel.com/seed…
193
Your next interior shoot might not need a photographer.
13 production pros, real photo vs AI, blind.
Interiors: real photo picked 52.8% of the time. Coin flip.
People: 80.6%. Food: 75.0%.
Rooms pass. Humans and food don't.
#GPTImage2 #Qwen #FLUX
journal.everypixel.com/which…
361
Did you realize right away that it’s an AI-generated picture?
What do you think—would it be possible to create that embroidery by hand?
ALT AI-generated picture of linen cloth with embroidery with flowers and herbs.
1
333
This seems very realistic too. We have many other beautiful pics at everypixel.com/free/exclusiv…
*Feel free to use these images in your commercial and personal projects.
182