gpustack/FLUX.1-lite-GGUF

!!! Experimental supported by gpustack/llama-box v0.0.84+ only !!!

Model creator: Freepik
Original model: flux.1-lite-8B-alpha
GGUF quantization: based on stable-diffusion.cpp ac54e that patched by llama-box.

Quantization	OpenAI CLIP ViT-L/14 Quantization	Google T5-xxl Quantization	VAE Quantization
FP16	FP16	FP16	FP16
Q8_0	FP16	Q8_0	FP16
(pure) Q8_0	Q8_0	Q8_0	FP16
Q4_1	FP16	Q8_0	FP16
Q4_0	FP16	Q8_0	FP16
(pure) Q4_0	Q4_0	Q4_0	FP16

---

Flux.1 Lite

We are thrilled to announce the alpha release of Flux.1 Lite, an 8B parameter transformer model distilled from the FLUX.1-dev model. This version uses 7 GB less RAM and runs 23% faster while maintaining the same precision (bfloat16) as the original model.

Text-to-Image

Flux.1 Lite is ready to unleash your creativity! For the best results, we strongly recommend using a guidance_scale of 3.5 and setting n_steps between 22 and 30.

import torch
from diffusers import FluxPipeline

base_model_id = "Freepik/flux.1-lite-8B-alpha"
torch_dtype = torch.bfloat16
device = "cuda"

# Load the pipe
model_id = "Freepik/flux.1-lite-8B-alpha"
pipe = FluxPipeline.from_pretrained(
    model_id, torch_dtype=torch_dtype
).to(device)

# Inference
prompt = "A close-up image of a green alien with fluorescent skin in the middle of a dark purple forest"

guidance_scale = 3.5  # Keep guidance_scale at 3.5
n_steps = 28
seed = 11

with torch.inference_mode():
    image = pipe(
        prompt=prompt,
        generator=torch.Generator(device="cpu").manual_seed(seed),
        num_inference_steps=n_steps,
        guidance_scale=guidance_scale,
        height=1024,
        width=1024,
    ).images[0]
image.save("output.png")

Motivation

Inspired by Ostris findings, we analyzed the mean squared error (MSE) between the input and output of each block to quantify their contribution to the final result, revealing significant variability.

As Ostris pointed out, not all blocks contribute equally. While skipping just one of the early MMDiT or late DiT blocks can significantly impact model performance, skipping any single block in between does not have a significant impact over the final image quality.

Future work

Stay tuned! Our goal is to distill FLUX.1-dev further until it can run smoothly on 24 GB consumer-grade GPU cards, maintaining its original precision (bfloat16), and running even faster, making high-quality AI models accessible to everyone.

ComfyUI

We've also crafted a ComfyUI workflow to make using Flux.1 Lite even more seamless! Find it in comfy/flux.1-lite_workflow.json.

The safetensors checkpoint is available here: flux.1-lite-8B-alpha.safetensors

HF spaces 🤗

You can also test the model on Flux.1 Lite HF space thanks to TheAwakenOne

Try it out at Freepik!

Our AI generator is now powered by Flux.1 Lite!

🔥 News 🔥

• Oct 28, 2024. Flux.1 Lite 8B Alpha HF space available on HF Space thanks to TheAwakenOne
• Oct 23, 2024. Alpha 8B checkpoint is publicly available on HuggingFace Repo.

Citation

If you find our work helpful, please cite it!

@article{flux1-lite,
  title={Flux.1 Lite: Distilling Flux1.dev for Efficient Text-to-Image Generation},
  author={Daniel Verdú, Javier Martín},
  email={dverdu@freepik.com, javier.martin@freepik.com},
  year={2024},
}

Attribution notice

The FLUX.1 [dev] Model is licensed by Black Forest Labs. Inc. under the FLUX.1 [dev] Non-Commercial License. Copyright Black Forest Labs. Inc.

Our model weights are released under the FLUX.1 [dev] Non-Commercial License.