Init

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+qwen3-4b-abl-q4_0.gguf filter=lfs diff=lfs merge=lfs -text

LANGUAGE_BIAS_STUDY.md

@@ -0,0 +1,138 @@
+---
+TL;DR: Flux.2 associates languages with styles:
+- Japanese=anime portraits
+- Chinese=anime scenes
+- German=illustrated art
+- English/Spanish=photorealistic
+---
+
+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<br>
+WANT ANIME FULL SCENES?         → Use Chinese<br>
+WANT ILLUSTRATED/STORYBOOK ART? → Use German<br>
+WANT PHOTOREALISTIC?            → Use English or Spanish<br>
+WANT ENCODER-DEPENDENT RESULTS? → Use Korean or Russian<br>
+
+
+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:<br>
+Cordux. (2026). Flux.2 Klein Language Bias Study. Hugging Face.<br>
+https://huggingface.co/Cordux/flux2-klein-4B-uncensored-text-encoder/blob/main/LANGUAGE_BIAS_STUDY.md

README.md

@@ -0,0 +1,59 @@
+---
+license: apache-2.0
+license_link: https://huggingface.co/Qwen/Qwen3-4B/blob/main/LICENSE
+base_model:
+- huihui-ai/Qwen3-4B-abliterated
+tags:
+- qwen3
+- abliterated
+- uncensored
+- text-generation-inference
+- flux2
+- klein
+extra_gated_prompt: >-
+  **This model has safety filtering removed and can generate General NSFW content.**
+  **By accessing this model, you agree to:**
+  - **Use it responsibly and legally**
+  - **Not use it to create illegal content**
+  - **Comply with all applicable laws in your country**
+
+---
+
+# Qwen3-4B Ablated (Uncensored) Text Encoder - GGUF Q4_0
+
+Uncensored/ablated version of Qwen3-4B text encoder in GGUF Q4_0 format for Flux2 Klein 4B models.
+
+## Compatible Models
+- Flux2 Klein 4B (Distilled & Base)
+
+## What This Does
+This is an ablated (safety-filtering removed) text encoder that allows Flux2 Klein models to generate NSFW content without prompt censorship.<br>
+The base Qwen3-4B text encoder that ships with Flux2 Klein has safety filtering that prevents certain prompts from being processed properly.
+
+## Installation
+1. Download `qwen3-4b-abl-q4_0.gguf`
+2. Place in `ComfyUI/models/text_encoders/` or `ComfyUI/models/unet/` (for GGUF loaders)
+3. In your workflow, use a GGUF-compatible text encoder loader node
+4. Point it to this file instead of the default Qwen3-4B
+
+## Prompting Tips
+- Use "wearing nothing" instead of "naked/nude" for best nude results
+- The model looks for clothing descriptors - even "nothing" counts as one
+- Clinical terms like "vagina" don't work better than colloquial terms
+- For explicit content beyond nudity, you'll need an NSFW LoRA
+
+### Language-Style Mapping Research
+I discovered Flux.2 Klein associates languages with specific styles: <br>
+Japanese→anime portraits, German→illustrated art, etc.<br>
+[Full study here](https://huggingface.co/Cordux/flux2-klein-4B-uncensored-text-encoder/blob/main/LANGUAGE_BIAS_STUDY.md)
+
+## Limitations
+This removes prompt filtering but doesn't add visual knowledge. The base Flux2 Klein models have limited training on explicit content, so:
+- ✅ Nudity works well
+- ✅ Suggestive poses work
+- ❌ Explicit anatomy requires a LoRA
+- ❌ Sexual acts require a LoRA
+
+## Credits
+- Based on [huihui-ai/Qwen3-4B-abliterated](https://huggingface.co/huihui-ai/Qwen3-4B-abliterated)
+- Converted with [llama.cpp](https://github.com/ggml-org/llama.cpp)

gitattributes

@@ -0,0 +1,37 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text
+qwen3-4b-abl-f16.gguf filter=lfs diff=lfs merge=lfs -text
+qwen3-4b-abl-q4_0.gguf filter=lfs diff=lfs merge=lfs -text

qwen3-4b-abl-q4_0.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:842d4ee7c62e8447b6bc122cb7616057c5da1333b0ef815b74812dc5db75047b
+size 2369547008