Flux2-Klein-9B-True-V2: Community Fine-Tune of FLUX.2-klein-9B Achieves Near-Nano-Banana Realism
Flux2-Klein-9B-True-V2 is the latest full fine-tune by wikeeyang based on FLUX.2-klein-9B, delivering significant improvements in photorealism, prompt adherence, LoRA compatibility, and image editing quality.
On July 1, 2026, community developer wikeeyang released Flux2-Klein-9B-True-V2, an updated full fine-tune of Black Forest Labs' FLUX.2-klein-9B. Building on the V1 release, V2 brings substantial gains in image realism, texture fidelity, prompt adherence, and LoRA compatibility — approaching the quality of Nano Banana-level outputs.
What is Flux2-Klein-9B-True-V2?
Flux2-Klein-9B-True-V2 is a fully fine-tuned checkpoint derived from FLUX.2-klein-9B, Black Forest Labs' compact 9B-parameter text-to-image model in the FLUX.2 family. As the successor to Flux2-Klein-9B-True-V1, it underwent a second round of full fine-tuning to push quality further.
The model has quickly gained traction in the community, amassing over 187,000 downloads and 161 likes on Hugging Face.
Key Improvements in V2
- Enhanced Photorealism & Texture: Image realism and material quality are significantly improved, now closely matching the standard set by Nano Banana fine-tunes.
- Better Prompt Adherence: The model follows complex prompts more faithfully and recovers fine details more accurately than V1.
- Improved LoRA Compatibility: Parameter adaptability and LoRA stack compatibility have been refined, making it easier to combine with community LoRAs.
- More Organized Structures: Image composition and structural integrity are more reasonable, with noticeable aesthetic improvements.
- Better Image Editing: Prompt adherence during image editing workflows has improved, with reduced color shift and pixel offset.
Supported Formats & Quantizations
The model is available in multiple formats for different hardware profiles:
| Format | File |
|---|---|
| BF16 full precision | Flux2-Klein-9B-True-v2-bf16.safetensors |
| FP8 mixed | Flux2-Klein-9B-True-v2-fp8mixed.safetensors |
| MX FP8 mixed | Flux2-Klein-9B-True-v2-mxfp8mixed.safetensors |
| NV FP4 mixed | Flux2-Klein-9B-True-v2-nvfp4mixed.safetensors |
| INT8 mixed row | Flux2-Klein-9B-True-V2-int8mixedrow.safetensors |
| GGUF Q4_K | Flux2-Klein-9B-True-v2-Q4_K.gguf |
| GGUF Q5_K | Flux2-Klein-9B-True-v2-Q5_K.gguf |
| GGUF Q6_K | Flux2-Klein-9B-True-v2-Q6_K.gguf |
| GGUF Q8_0 | Flux2-Klein-9B-True-v2-Q8_0.gguf |
Acceleration with LoRA
The model supports Turbo and Distill LoRAs for faster inference:
- Turbo LoRA by anyMODE (rank 64/128 on Civitai / HF): use weight 0.2–0.3 for 4–8 steps.
- Distill LoRA by vafipas663 (rank 128/256 on HF): use weight 0.4–0.6 for 4–8 steps.
For color and shape correction in edited images, the ComfyUI-LCS plugin is recommended for latent correction.
Usage Notes
The model uses the same workflow and settings as V1. It is also available on ModelScope. The provided edit workflow file (V2-Edit-with-LCS-example-workflow.json) can help users get started with image editing tasks.
License
This model falls under the FLUX Non-Commercial License from Black Forest Labs, and is intended for non-commercial research and community use.