Flux SVDQuants
Collection
8 items • Updated • 8
import torch
from diffusers import DiffusionPipeline
# switch to "mps" for apple devices
pipe = DiffusionPipeline.from_pretrained("spooknik/PixelWave-SVDQ", dtype=torch.bfloat16, device_map="cuda")
prompt = "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k"
image = pipe(prompt).images[0]This repository contains Nunchaku-quantized (SVDQ) versions of PixelWave, a text-to-image model based on Flux.1 Dev by humblemikey
Model Files
- svdq-int4_r32-pixelwave-schnell-04.safetensors: SVDQuant INT4 (rank 32) PixelWave FLUX.1-schnell 04. For users with non-Blackwell GPUs (pre-50-series).
- svdq-fp4_r32-pixelwave-schnell-04.safetensors: SVDQuant NVFP4 (rank 32) PixelWave FLUX.1-schnell 04. For users with Blackwell GPUs (50-series).
- svdq-int4_r32-pixelwave-dev-0.3.safetensors: SVDQuant INT4 (rank 32) PixelWave FLUX.1-dev 03. For users with non-Blackwell GPUs (pre-50-series).
- svdq-fp4_r32-pixelwave-dev-0.3.safetensors: SVDQuant NVFP4 (rank 32) PixelWave FLUX.1-dev 03. For users with Blackwell GPUs (50-series).
Quality Evaluation
Below is the quality and similarity evaluated with 256 samples from MJHQ-30K dataset. (BF16 is the unqauntized model. INT W4A4 is INT4 and NVFP4 is FP4)
| Model | Precision | Method | FID | IR | LPIPS | PSNR |
|---|---|---|---|---|---|---|
| PixelWave schnell 04 | BF16 | -- | 176.37 | 0.813 | -- | -- |
| (8 step) | INT W4A4 | SVDQ | 176.68 | 0.820 | 0.322 | 17.12 |
| NVFP4 | SVDQ | 176.91 | 0.839 | 0.298 | 17.70 |
If you find these useful please consider: 
- Downloads last month
- 54
Model tree for spooknik/PixelWave-SVDQ
Base model
black-forest-labs/FLUX.1-schnell