The "One House Look" Strategy Finished Last
We took the same source prompts, ran them through many different rewriting strategies, and generated thousands of images.
Two independent evaluators then ranked the results in pairwise comparisons.
What separated the winners from the losers was judgment, not vocabulary:
・Deciding where the viewer's eye should land
・Choosing light with a visible source
・Letting composition follow the kind of image being made
The strategy that stamped one fixed stylistic treatment onto every prompt came last on both evaluators.
That is our clearest evidence against giving PixAI a house look.
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.
People have been asking how we made this so here are the tools that we used to make the video:
Character Design: @PixAI_Official
Prompt Control: Atelier Magic (In House)
Storyboard: Atelier Boards (In House)
Video Gen: Seedance 2.5 @get_artcraft
Check it out. An AI Malfoid Contest on PixAI! 🐍✨
Submit your Malfoid AI art to the contest: https://t.co/07y54Ifcxn
Want to rally more likes on your entry? Retweet this post with your submitted image and your PixAI post link.
Deadline: 10/24/2026
#PixAI#Malfoid@PixAI_Official
Tsubaki.3 Technical Report
https://t.co/wWy8e9bQ6P
Tsubaki.3 is our multimodal foundation model for anime image generation, spanning illustration, manga, webtoon, and more. The report walks through its data curation, training strategy, post-training curriculum, and evaluation.
We’ve also open sourced Tagger 1.0, our anime image tagging model, to support further work by the community. Tagger 1.0 on Hugging Face: https://t.co/qtVIN2k1av
#PixAITsubakiV3#PixAIPrompt
Tsubaki.3で使えるかもしれないプロンプト
流行りの『"素人の失敗写真" 3×3グリッド』
同じキャラ・同じ衣装のブレ写真・逆光・指写り込み…を9コマまとめて出せます👇(↓生成例)
Panel欄は好きな内容で埋めてOK!!
Composition:anime coloring, cg, no overlapping panels, thin black borders, all elements inside panels, no bleed, 9 even panels starting from top-left to bottom-right in a grid formation depicting this character in a consistent outfit.
Character:
Panel 1 (top-left):
Panel 2 (top-middle):
Panel 3 (top-right):
Panel 4 (center-left):
Panel 5 (center-middle):
Panel 6 (center-right):
Panel 7 (bottom-left):
Panel 8 (bottom-middle):
Panel 9 (bottom-right):
試したら引用リポストで生成結果を見せてください🙌