AEmotionStudio/flux-klein-fp8

FLUX Klein 4B FP8 — AEmotionStudio Mirror

FP8-quantized mirror of black-forest-labs/FLUX.2-klein-4B for use with ComfyUI-FFMPEGA.

See also: AEmotionStudio/flux-klein (full bf16 weights)

About

FLUX Klein is a 4-billion parameter image editing and generation model from Black Forest Labs. It supports:

• Object removal — erase unwanted objects from images/video frames
• Text-guided editing — change hair color, swap backgrounds, modify clothing, etc.
• Inpainting — fill masked regions with contextually appropriate content

This repo provides FP8-quantized (float8_e4m3fn) weights, reducing download size and VRAM usage by roughly half compared to the bf16 mirror with negligible quality loss for most editing tasks.

Model Details

Property	Value
Architecture	FLUX (Rectified Flow Transformer)
Parameters	4B
Precision	fp8 (float8_e4m3fn)
Total Size	~8 GB
License	Apache 2.0
Source	black-forest-labs/FLUX.2-klein-4B
bf16 Mirror	AEmotionStudio/flux-klein

Files

Component	File	Size
Transformer	transformer/diffusion_pytorch_model.safetensors	3.88 GB
Text Encoder (T5-XXL)	text_encoder/model.safetensors	4.02 GB
Configs	model_index.json, transformer/config.json, text_encoder/*.json	< 60 KB

Note: VAE, tokenizer, and scheduler configs are not included in this repo. They are loaded from the upstream pipeline defaults at runtime.

Usage with ComfyUI-FFMPEGA

This model is auto-downloaded when you select flux_klein_precision = "fp8" in ComfyUI-FFMPEGA and use the auto_mask skill with effect=remove or effect=edit.

Object Removal

Prompt the FFMPEGA Agent:

Remove the watermark

Or use the Effects Builder with SAM3 target + remove effect.

Text-Guided Editing

Prompt the FFMPEGA Agent:

Change the hair to red

Or use auto_mask with effect=edit and edit_prompt="change the hair to red".

Manual Download

If auto-download is disabled, download all files and place them in:

ComfyUI/models/flux_klein_fp8/

VRAM Requirements

• Minimum: ~6 GB (with aggressive CPU offloading)
• Recommended: ~8 GB (fp8, sequential CPU offload)
• Roughly half the VRAM of the bf16 variant, making it practical for 8 GB consumer GPUs (RTX 3060/4060)

FP8 vs BF16

	FP8 (this repo)	BF16 (flux-klein)
Download	~8 GB	~15 GB
VRAM	~8 GB rec.	~13 GB rec.
Quality	Negligible loss	Full precision
Speed	Faster on Ampere+	Baseline

License

The model weights are licensed under Apache 2.0 by Black Forest Labs. See LICENSE.md for full terms.

Citation

@misc{flux2klein2025,
  title={FLUX.2 Klein},
  author={Black Forest Labs},
  year={2025},
  url={https://huggingface.co/black-forest-labs/FLUX.2-klein-4B}
}