DiffSynth-Studio/flux-3-image-klein-entropack

FLUX.3 Image Klein — EntroPack Pre-Quantized Models

We provide pre-quantized weight packages produced by EntroPack for FLUX.3 Image Klein (black-forest-labs/flux-3-image-klein), which can be used with the quantization features of DiffSynth-Studio.

Both the DiT and the text encoder are provided at three compression rates (3 / 4 / 6 bit), and the DiT additionally ships FP8 GEMM variants:

• 3bit (lowest VRAM) — mixed-bit: the most sensitive layers (embeddings, time / modulation and final layers, 15 layers) keep 4.0 bpp, while the remaining 258 layers use 2.9482 bpp (weighted target 3.0 bpp);
• 4bit (balanced) — uniform entropy coding at 4 bits per weight;
• 6bit (highest precision) — uniform entropy coding at 6 bits per weight; the accuracy is close to FP8 quantization;
• fp8_3bit / fp8_4bit / fp8_6bit (DiT only) — weights are first FP8 W8A8-quantized, then entropy-coded by EntroPack down to 3 / 4 / 6-bit storage; the forward pass uses FP8 GEMM for faster inference.

All packages are entropy-coded from the original BF16 weights directly. Every *.safetensors ships with a same-named *.json that records its quantization config. actual bpp = 8 × Σ storage bytes / Σ numel is measured on the package artifacts.

DiT

name	file	actual bpp	size	preview
dit_3bit (mixed-bit)	models/dit_3bit.safetensors	3.021	2.10 GB
dit_4bit	models/dit_4bit.safetensors	4.024	2.79 GB
dit_6bit	models/dit_6bit.safetensors	6.053	4.20 GB
dit_fp8_3bit (FP8 GEMM, mixed-bit)	models/dit_fp8_3bit.safetensors	3.030	2.10 GB
dit_fp8_4bit (FP8 GEMM)	models/dit_fp8_4bit.safetensors	4.032	2.80 GB
dit_fp8_6bit (FP8 GEMM)	models/dit_fp8_6bit.safetensors	6.049	4.20 GB

Text Encoder

name	file	actual bpp	size
text_encoder_3bit	models/text_encoder_3bit.safetensors	3.023	2.46 GB
text_encoder_4bit	models/text_encoder_4bit.safetensors	4.027	3.02 GB
text_encoder_6bit	models/text_encoder_6bit.safetensors	6.058	4.14 GB

Usage

Install

# DiffSynth-Studio
git clone https://github.com/modelscope/DiffSynth-Studio.git
cd DiffSynth-Studio && pip install -e .

# EntroPack (pick the extra matching your CUDA version: cuda13 / cuda12)
git clone https://github.com/modelscope/entropack.git
cd entropack && pip install -e '.[cuda13]'

Example scripts

• Inference: standard / low VRAM
• LoRA training: train / validation

dit_package and text_encoder_package are chosen independently: any combination works (e.g. dit_fp8_6bit with text_encoder_4bit). dit_4bit and text_encoder_4bit are pre-registered in DiffSynth-Studio, so they load with a plain ModelConfig; the helper below registers any package's json sidecar before loading:

import json

from diffsynth.configs import MODEL_CONFIGS
from diffsynth.pipelines.flux3_image import Flux3ImagePipeline, ModelConfig

prequant_model_id = "DiffSynth-Studio/flux-3-image-klein-entropack"
dit_package = "models/dit_4bit.safetensors"
text_encoder_package = "models/text_encoder_4bit.safetensors"


def make_package(pattern):
    package = ModelConfig(model_id=prequant_model_id, origin_file_pattern=pattern)
    package.download_if_necessary()
    sidecar = ModelConfig(model_id=prequant_model_id, origin_file_pattern=pattern.replace(".safetensors", ".json"))
    sidecar.download_if_necessary()
    with open(sidecar.path) as handle:
        entry = json.load(handle)
    if not any(config.get("model_hash") == entry["model_hash"] for config in MODEL_CONFIGS):
        MODEL_CONFIGS.append(entry)
    return package


pipe = Flux3ImagePipeline.from_pretrained(
    torch_dtype=torch.bfloat16,
    device="cuda",
    model_configs=[
        make_package(dit_package),
        make_package(text_encoder_package),
        ModelConfig(model_id="black-forest-labs/flux-3-image-klein", origin_file_pattern="vae/diffusion_pytorch_model.safetensors"),
    ],
    processor_config=ModelConfig(model_id="black-forest-labs/flux-3-image-klein", origin_file_pattern="text_encoder/"),
)