[MV Release] PixAI’s 4th Anniversary Special Arrangement of “The Colors Unseen” is now live! 🎁 Win a share of $500 in Amazon Gift Cards ($50 × 10 winners) Follow @PixAI_EN, then like and repost this post to enter! ※10 winners will be selected across both the JP and global official accounts. ▼ Song Credits Title: The Colors Unseen Vocals: chelly（チェリー）@chelly__L51O93 Lyrics: 山崎寛子／Hiroko Yamazaki Composition: 神谷礼／Rei Kamiya Arrangement：菊谷知樹／Tomoki Kikuya Music Production: POPHOLIC Animation: PixAI Animation Team ▼ YouTube youtu.be/VcQRYLfLlYo

Style Names Plateau at 60% During pretraining we tracked Recall@5 on frozen rosters of concepts. ・Character names: reached 95.8% ・General concepts: reached 91.5% ・Style proper nouns: climbed from 15.6% to 60.6%, then flattened near 60% for the rest of the run Under the natural data distribution, rare style concepts simply do not appear often enough. Training longer does not close the gap. That plateau is exactly why the long-tail oversampling from Day 9 exists.

Two personalities. One bean bag. You can guess who won.

An Open Problem: Better at Instructions, Slightly Worse at Style We score in-context generation with a VLM against human-written checklists. Its answers agree with human judges 85% of the time. Across three checkpoints, C1 → C2 → C3, the checklist pass rate goes up. At the same time, the style-transfer score goes down, and the style-preservation score drifts further from zero. In plain terms: the model follows instructions better, and matches the reference's art style slightly worse. This is not solved yet. The figure is in the report exactly as it is.

📢 Login Issue Resolved Earlier today, some users may have unexpectedly been logged out of their PixAI accounts, and this was due to a known server side issue that has now been resolved. If you were unexpectedly logged out, please restart the app or browser and simply log in again. If you find yourself logged into a different account, this may be because Google login and email login are treated as separate accounts — please try logging in with the method you previously used. If you continue to experience issues logging back in, please contact our support team or reach out to a moderator in our Discord community. We apologize for the inconvenience and thank you for your patience. The PixAI Team

5,000 of you?! Thank you for making this account what it is. Now we want to see what you've been making. Quote this post with your favorite PixAI creation from the past few weeks. How to Join 1. Follow @PixAI_EN 2. Quote this post with your artwork from PixAI + @PixAI_EN Submission Period Oct 17, 2026 – Oct 24, 2026, 23:59 (UTC+9) Prizes • $50 Amazon Gift Card × 2 winners • 500,000 PixAI Credits × 10 winners Random draw, one entry per person.

[PixAI 4th Anniversary] PixAI 4th Anniversary: Project Genesis | 3 Days Left! Only 3 days left! • 50% OFF Annual PREMIUM Membership + Tsubaki.3 Unlimited Mode • Up to 100% Bonus Credits on credit packs Join the celebration: • "UnLocking World" MV Giveaway & Creative Challenge: $3,000 in Amazon Gift Cards + a chance to win a signed shikishi • Tsubaki.3 Creation Contest: 80M credits + $3,000 in Amazon Gift Cards • "The Colors Unseen" MV Giveaway: $500 in Amazon Gift Cards • Create Nasu-themed artworks or share PixAI offline ad photos ⏰ Ends Oct 20, 2026, at 12:00 (UTC+9) pixai.art/en/event/4th-anniv…

Choosing the Final Checkpoint: Aesthetics vs. Correctness Every candidate checkpoint is placed on the same two axes. ・X axis: image quality and aesthetics ・Y axis: concept coverage and correctness Both are scored by in-house models. The four colored points in the figure trace the whole post-training path. Pretraining sets the starting point. SFT raises quality and slightly lowers correctness. Correctness RL lifts correctness to the highest value on the chart, at some cost in quality. Aesthetic RL recovers quality while correctness stays near the top. The point labeled Tsubaki.3 is where we chose to sit on that frontier.

[Winners annunced] PixAI Tsubaki.3 creation event Congratulations to our 10 winners! Each of you will receive 200,000 PixAI Credits 🎁 @Tokoyo4471 @shirasu561 @HailyM53661 @makuya0526 @SakoSakokiti @Matasanos09 @Lumina_1023 @_user1003_ @kantakanta1233 @daisuke41617 How to Claim Please DM @PixAI_EN with your PixAI profile link. We'll verify each claim and add the credits in order after confirmation. Thank you to everyone who joined!

PixAI × zeta OC Creation Contest 2026 is coming in one week! Your next adventure with your OCs starts here. Generate your OC on PixAI. Build their world and plot on zeta. Step into their story. Event period: Oct 23 – Nov 4 (KST) 🎁 180,000,000 PixAI Credits + 359,000 zeta Pieces Share the event page on X and get 5,000 PixAI Credits to start creating! ▼ Event page ▼ pixai.art/en/event/zeta

Enter the contest: pixai.art/en/contest/user-ma… Watch the stream: twitch.tv/chaotichelleaven More about Chaotic Helleaven: linktr.ee/chaotichelleaven

Chaotic Helleaven × PixAI: Malfoid Art Reaction Stream! 🐍 Your MalfoidFever entries are about to get a bigger audience. Divanyaa, the creator, writer and editor of Chaotic Helleaven (@CHelleaven), is going live on Twitch to react to entries from PixAI's Malfoid Fever contest. • Where: twitch.tv/chaotichelleaven • When: October 15, 2026 · 4:00 PM ET / 1:00 PM PT Who's Chaotic Helleaven? A comedy media franchise about Rika Helleaven, an overpowered otaku who is the daughter of God and the Devil. Heaven and Hell both need her on their side before the apocalypse can start. It has manga, community-funded animations and VTuber-style streams, and the humor is as chaotic as it sounds. Haven't entered yet? There's still time! The Malfoid Fever contest is hosted by our community member differentwarning. • Prize pool: 500,000 PixAI credits • Deadline: October 24, 2026 • How to enter: publish your artwork on PixAI with #MalfoidFever Contest and stream links in the reply 👇

Reward Hacking: Every Dip Is a Correction, Not a Restart The hardest problem in RL is reward hacking. Training reward and validation scores keep climbing while the actual images get worse. Two defenses. ・Detection: a fixed, hand-picked evaluation set, generated with fixed seeds at every checkpoint. It is chosen for sensitivity to known failures: washed-out overexposure, synthetic grain, distorted lettering. When scores rise but the eval set degrades, training stops ・Recovery: a short SFT pass pulls the stopped checkpoint back toward real data. Then RL resumes and climbs past the old peak In the reward curve, every dip is one of these corrections. None of them is a restart.

This is amazing. The character design was made with PixAI by @Atelier452

900k watching. 0 comments. This is the most peaceful stream I've ever done 🚀

One Rubric per Prompt, Written by an Agent Many projects use a vision-language model as the reward for RL. But a single unified reward cannot serve every image. What is critical for one prompt is trivial for another. In our early experiments, one shared standard actually accelerated reward hacking. So we built an agent that writes a rubric for each prompt. Six tools, a four-stage loop: ・Draft: write the rubric for this prompt ・Recalibrate: adjust the verifier only if needed. Any edit must pass a human-labeled anchor set before it is frozen again ・Probe: test the rubric on outputs from the current policy ・Prune and backfill: drop prompts where everything passes or everything fails, and add new ones Every rubric is readable and auditable by a human.

Emoji Challenge

[PixAI 4th Anniversary] Project Genesis Special Event | 7 Days Left! The final countdown is on! Don't miss your last chance to enjoy anniversary perks and join our special events. · 50% off annual PREMIUM memberships, up to 100% bonus credits, and special Tsubaki.3 video offers! · Collect Mio gacha cards, enjoy free daily generation, join the Tsubaki.3 Creation Contest, and take part in our MV and Nasu events for a chance to win exciting rewards! ⏰ Ends October 20, 2026, at 12:00 (UTC+9) 👉 Explore the Anniversary Events: pixai.art/en/event/4th-anniv…

Two-Stage RL: Correctness First, Aesthetics Second Reinforcement learning runs as a two-stage cascade. Stage 1 is correctness. Four tasks, each with its own prompt pool and its own reward: ・Anatomy ・Multi-subject scenes ・Character interaction ・Text rendering At every training step, one task is drawn at random. Two tasks never share a batch, so a strong gradient from one task cannot wash out a weaker one. In the figure, left is after SFT and right is after Stage 1. The malformed hand, the extra arm, the twisted table legs: repaired. They also read a little stiffer. Stage 2, aesthetic RL, pays that back.