Krea 2 Turbo W4A4 (Nunchaku)
Krea 2 Turbo quantized to 4-bit weights and 4-bit activations with SVDQuant, running through Nunchaku's fused low-bit kernels. As far as I can tell this is the first W4A4 build of Krea 2, and the first time it has generated an image through fused 4-bit kernels rather than a simulation of them.
This is not storage quantization. The fp8 and int8 builds in circulation shrink the file and leave the forward-pass arithmetic untouched. This one rewrites the arithmetic, which is the only kind of quantization that buys real time.
Measured
Scored against BF16 on the same card, 12 versioned prompts, 8 fixed seeds, 96 seed-locked pairs per resolution. Latency is a warm median of 3 seeds on an L40S with both models on the same attention path.
| fidelity vs bf16 | LPIPS | PSNR | ImageReward delta | n |
|---|---|---|---|---|
| 1024px | 0.276 | 17.7 | +0.078 | 96 |
| 512px | 0.226 | 18.4 | -0.013 | 96 |
| latency | W4A4 | bf16 | speedup |
|---|---|---|---|
| 1024px, 8 steps | 5.36s | 7.71s | 1.44x |
| 512px, 8 steps | 1.34s | 2.19s | 1.63x |
| 512px, 2 steps | 0.41s | n/a | 2.45 img/s |
Quality did not degrade. A positive ImageReward delta is not a quality gain and I would rather say that plainly than let the number flatter the build. A reward model is a learned preference with known failure modes, and the likelier reading is that quantization nudged the trajectory into a more prototypical basin. What the number supports is the narrower claim: four bits of weight and four bits of activation, and no penalty this measure can see. The 1024px result lands on the fake-quant study's prediction of 0.268, which is the best evidence that the running build is faithful to the calibration it came from.
The reason it holds is the shape of the method. SVDQuant carries a rank-32 full-precision branch next to the 4-bit matmul, and it absorbs the outlier activations that would otherwise wreck the last blocks. Calibration error climbs roughly 3x from the first block to the last, and the branch soaks it up.
Read this before you benchmark it
Krea 2 uses grouped-query attention and always passes an attention mask, because text and
image share one sequence. PyTorch's SDPA will not serve enable_gqa=True together with a
mask on the flash backend. It drops to the math backend silently and materializes the full
attention matrix. Expanding the 12 k/v heads to 48 by hand restores FlashAttention with the
mask still in place.
Four-way ablation on the real pipeline, 1024px, 8 steps:
| time | |
|---|---|
enable_gqa=True |
22.3s |
enable_gqa=True + dtype cast |
22.4s |
| expanded heads | 7.7s |
| expanded heads + dtype cast | 7.8s |
A separate 3x, sitting on top of both columns above. Any model that splits heads this way and also masks will lose the fast kernel without saying so. I got the cause wrong twice before isolating it, once by testing the attention call outside the pipeline where there was no mask to trigger the condition.
Usage
Requires the Krea 2 runtime for Nunchaku. It is under review upstream; until it lands, the files are at github.com/sztlink/krea2-nunchaku.
from nunchaku.models.transformers.transformer_krea2 import NunchakuKrea2Transformer2DModel
from diffusers import Krea2Pipeline
import torch
transformer = NunchakuKrea2Transformer2DModel.from_pretrained(
"felipesztutman/Krea-2-Turbo-W4A4-Nunchaku/svdq-int4_r32-krea-2-turbo.safetensors",
torch_dtype=torch.bfloat16,
)
pipe = Krea2Pipeline.from_pretrained(
"krea/Krea-2-Turbo", transformer=transformer, torch_dtype=torch.bfloat16,
).to("cuda")
# Krea 2 Turbo is cfg-distilled. guidance_scale MUST be 0.0.
image = pipe("a fox in the snow", guidance_scale=0.0, num_inference_steps=8).images[0]
Ampere or newer, verified on an RTX 3090 (sm_86) on both Linux and native Windows 11: 2.7s at
512px/8 steps, 9.5s at 1024px/8 steps, 0.85s at 512px/2 steps, warm, clean image on both. Also
runs on Ada (L40S). Use nunchaku==1.2.1 (cu12.8 torch2.8 cp312); it carries the loader this
runtime needs and ships both Linux and Windows wheels. Nunchaku reports INT4 support down to
Turing since v1.2.0, untested here. The NVFP4 variant needs Blackwell; this INT4 build does not.
See the runtime repo for the exact install.
Provenance and limits
Calibrated on an H100 with deepcompressor, 128 calibration prompts. Every fidelity number above traces to raw per-pair data in dead-channel, the open benchmark this came out of, including the pairs where it does worst.
Known: compositions shift relative to BF16 at the same seed, which is expected and is what LPIPS 0.276 is reporting. Not a drop-in seed-for-seed replacement if you need to reproduce existing BF16 outputs exactly.
License and attribution
This is a modified version of Krea 2 Turbo. It is not official, and it is not endorsed by
or affiliated with Krea. Use is governed by the Krea 2 Community License Agreement, a copy of
which is included in this repository as LICENSE.pdf.
Krea 2 is licensed under the Krea 2 Community License Agreement. For more information, visit https://krea.ai/krea-2-licensing.
The base model ships with a content filter. Quantization does not remove or weaken it, and redistributing this checkpoint does not license you to bypass it. Downstream use remains subject to the base license's acceptable use terms.