LANGUAGE_BIAS_STUDY.md

FLUX.2 KLEIN 4B - LANGUAGE TO STYLE MAPPING RESEARCH Complete Findings - January 27, 2026

TEST CONDITIONS:

• Model: Flux.2 Klein 4B Distilled
• Base Parameters: 4 steps, CFG 1.0, 1024x1024 resolution
• Text Encoders Tested: Ablated Qwen3-4B GGUF vs Original FP4 Encoder
• Test Prompt: "Red-haired woman, sitting on chair, one hand holding fireball, other hand holding lightning" (translated to each language)

COMPUTER:

• NVidia RTX 3050 6GB VRAM
• 32 GB RAM
• Intel(R) Core(TM) i3-9100F COU @ 3.6GHz
• WS Blue SN570 1TB NVMe Drive

TOOLS USED:

• Model Execution: ComfyUI with custom workflows (Built myself)
• Text Encoders: Ablated Qwen3-4B GGUF, Original FP4
• Note Management: Notepad
• Analysis: Manual comparison of 14+ generated images / rgthree Image Compare Node (For direct compare the images)

COMPLETE RESULTS TABLE

LANGUAGE	ABLATED QWEN3 ENCODER	ORIGINAL FP4 ENCODER	CONCLUSION
English	Photorealistic	Photorealistic	Neutral - Consistent realism
Japanese	Anime	Anime (character focused)	STRONG inherent anime bias
Chinese	Anime	Anime	STRONG inherent anime bias
Korean	Anime	Realistic	Ablation-induced bias only
German	Illustrated/Painted	Illustrated/Painted	STRONG inherent art bias
Russian	Semi-realistic/Fantasy	Realistic	Ablation-induced bias only
Spanish	Photorealistic	Realistic (+fireball detail)	Neutral with detail variation

KEY DISCOVERIES:

1. INHERENT TRAINING DATA BIASES (Present in both encoders):

   • Japanese → Anime character portraits (strongest association)
   • Chinese → Anime full scenes
   • German → Illustrated storybook/fairy tale art
   • English/Spanish → Photorealistic (neutral baseline)

2. ABLATION-INDUCED BIASES (Only in ablated encoder):

   • Korean → Anime style (not inherent)
   • Russian → Fantasy hybrid style (not inherent)

3. JAPANESE vs CHINESE NUANCE:

   • Japanese prompts focus on CHARACTER (often no background)
   • Chinese prompts include FULL SCENES with simple backgrounds
   • Suggests different dataset types for each language

PRACTICAL STYLE GUIDE:

For CONSISTENT RESULTS, use these language choices:

WANT ANIME CHARACTER PORTRAITS? → Use Japanese
WANT ANIME FULL SCENES? → Use Chinese
WANT ILLUSTRATED/STORYBOOK ART? → Use German
WANT PHOTOREALISTIC? → Use English or Spanish
WANT ENCODER-DEPENDENT RESULTS? → Use Korean or Russian

ENCODER RECOMMENDATIONS:

ORIGINAL FP4 ENCODER:

• More consistent across languages
• Reveals true Flux.2 training biases
• Recommended for predictable style control

ABLATED QWEN3 ENCODER:

• Amplifies all style associations
• Adds extra biases (Korean→anime, Russian→fantasy)
• Use for exaggerated stylistic effects

TRAINING DATA INFERENCES:

Based on these results, Flux.2 was likely trained on:

1. Japanese datasets = Anime character sheets/portraits
2. Chinese datasets = Anime full scenes/webtoons
3. German datasets = Illustrated books/fairy tale art
4. English/Spanish datasets = Photographic stock images
5. Korean/Russian datasets = Mixed/general content

RESEARCH NOTES:

• Korean losing anime bias with original encoder suggests Korean training data was more general-purpose than Japanese/Chinese anime-specific data.

• Spanish showing increased fireball detail attention suggests subtle language-specific attention patterns beyond just style.

• The character-focused vs scene-focused difference between Japanese and Chinese outputs indicates the model distinguishes between different types of anime content by language origin.

NEXT RESEARCH QUESTIONS:

1. Do these biases exist in full Flux.2-dev (12B) or only Klein (4B)?
2. How does quantization (Q4 vs FP4 vs FP16) affect bias strength?
3. Can prompt engineering override language biases? (e.g., "Photorealistic: [Japanese text]" or "Anime style: [German text]")
4. Do other multilingual encoders (CLIP, T5) show similar patterns?

SPECIAL NOTE:

Methodological Transparency:

• Research design & testing: Human (Cordux)
• Data analysis & conclusions: Human (Cordux)
• Documentation formatting, Keeping track of notes & clarity: AI-assisted (DeepSeek, Claude)

END OF RESEARCH DOCUMENT

Citation:
Cordux. (2026). Flux.2 Klein Language Bias Study. Hugging Face.
https://huggingface.co/Cordux/flux2-klein-4B-uncensored-text-encoder/blob/main/LANGUAGE_BIAS_STUDY.md