FLUX.1 Kontext Dev Nunchaku Lite INT4 r32
Diffusers-loadable conversion of:
- Base model:
black-forest-labs/FLUX.1-Kontext-dev - Source repo:
nunchaku-ai/nunchaku-flux.1-kontext-dev - Source checkpoint:
svdq-int4_r32-flux.1-kontext-dev.safetensors
The CLIP text encoder is copied from the base model, and the T5 text_encoder_2 is saved with BitsAndBytes 4-bit NF4.
Benchmark
| Checkpoint | Latency | Max VRAM |
|---|---|---|
| Converted Diffusers Nunchaku Lite INT4 r32 + BNB4 T5 encoder | 32.89 s | 13.85 GiB |
Output Comparison
The comparison uses the same input image, prompt, seed, scheduler, and 28-step settings for this converted checkpoint and its original Nunchaku safetensors baseline. The converted checkpoint runs through the stock Diffusers graph with the Nunchaku Lite quantizer — no fused-GELU patches and no split-linear patches.
Run
Requires a Turing, Ampere, Ada, or Blackwell NVIDIA GPU (INT4 kernels are not supported on Hopper) and the Hugging Face kernels package.
import torch
from PIL import Image
from diffusers import FluxKontextPipeline
model_path = "lite-infer/flux.1-kontext-dev-nunchaku-lite-int4_r32-bnb4-text-encoder"
image_path = "input.png"
output_path = "output.png"
pipe = FluxKontextPipeline.from_pretrained(model_path, torch_dtype=torch.bfloat16)
pipe.to("cuda")
image = Image.open(image_path).convert("RGB")
prompt = "Make Pikachu hold a sign that says 'Nunchaku is awesome', yarn art style, detailed, vibrant colors"
result = pipe(
image=image,
prompt=prompt,
generator=torch.Generator(device="cuda").manual_seed(1),
true_cfg_scale=1.0,
guidance_scale=2.5,
num_inference_steps=28,
max_area=512 * 512,
)
result.images[0].save(output_path)