You can deploy this model and start earning money today!

Owen777/
flux.1-lite-8B-GRPO

Name: flux.1-lite-8B-GRPO
Availability: InStock

Freepik/flux.1-lite-8B

finetune

apache-2.0

Views

No views yet

flux.1-lite-8B-GRPO (DanceGRPO + PickScore)

A reward-aligned transformer for Flux 1‑Lite 8B, fine-tuned with DanceGRPO using PickScore as the reward signal.
This repository hosts the FluxTransformer2DModel weights only; load them on top of the original Freepik/flux.1-lite-8B base pipeline.

Highlights

🔁 RL post-training (DanceGRPO) for preference alignment

🏆 PickScore as the reward model to boost aesthetics and prompt faithfulness

⚡ Drop-in replacement for the base model’s transformer in Diffusers

Quick Start (Diffusers)

python

import torch
from PIL import Image
from diffusers.pipelines.flux.pipeline_flux import FluxPipeline
from diffusers.models.transformers.transformer_flux import FluxTransformer2DModel

# 1) Load the base Flux 1-Lite pipeline
base_model = "Freepik/flux.1-lite-8B"
pipe = FluxPipeline.from_pretrained(base_model, torch_dtype=torch.bfloat16).to("cuda")

# 2) Swap in the GRPO-tuned transformer from this repo
grpo_transformer = "Owen777/flux.1-lite-8B-GRPO"
transformer = FluxTransformer2DModel.from_pretrained(grpo_transformer, torch_dtype=torch.bfloat16).to("cuda")
pipe.transformer = transformer  # replace in-place


prompt="A studio portrait of a woman with bold makeup, highly detailed eyelashes, and glossy red lips, dramatic directional lighting creating high contrast shadows, perfect catchlights in the eyes, smooth skin texture without blemishes"
image = pipeline(prompt, 
    num_inference_steps=50, 
    guidance_scale=2.5,
    height=1024,
    width=1024,
    generator=torch.Generator(device="cuda").manual_seed(42)
    ).images[0]
image.save("flux_1_lite_grpo.png")

Tips

Start with num_inference_steps=28–60, guidance_scale=2.5–4.5.
Use BF16 on modern NVIDIA GPUs (A100/H100/H200/4090).
For strict reproducibility, set a fixed generator seed.

What’s in this repo?

Weights: FluxTransformer2DModel only (no VAE, no tokenizer, no text encoder).
How to use: Instantiate the base Flux 1‑Lite 8B pipeline and replace its transformer with ours (see Quick Start).

Training Summary

Base model: Freepik/flux.1-lite-8B
Algorithm: DanceGRPO (a GRPO-style preference optimization method)
Reward model: PickScore (image–text aesthetic & preference scoring)

Objective (high level)

We frame post-training as preference optimization: for a batch of prompts, multiple candidates are generated and scored by PickScore. DanceGRPO then updates the policy (the transformer) to increase the relative likelihood of higher-scoring samples while decreasing it for lower-scoring ones. This tends to:

Improve aesthetic quality favored by PickScore,
Encourage closer adherence to prompt intent, and
Stabilize detail rendition at higher resolutions.

Note: Dataset specifics, compute budget, and exact hyperparameters are not disclosed here. The method is model-agnostic and can be applied to other Flux variants.

Inference Recommendations

Resolution: 768–1024 on the long side is a good default.
Guidance: Lower guidance can look more natural; higher guidance is more literal but may reduce diversity.
Seeds: Fix a seed for deterministic results; sweep seeds for variety.
Batching: Use smaller batches when memory is tight; enable torch.backends.cuda.matmul.allow_tf32 = True if your environment supports it.

Limitations & Notes

Reward alignment reflects what PickScore prefers; it may over-optimize for aesthetics in some edge cases.
The model may still struggle with complex spatial compositing, tiny text, or very long prompts.
Safety and content filtering should be handled upstream (prompting) and/or downstream (moderation) according to your use case.

Compatibility

Diffusers Flux pipeline with transformer swap-in (as shown above).
Torch dtype: BF16 preferred on modern NVIDIA GPUs.
OS/Env: Standard PyTorch CUDA environments; no custom CUDA ops are required for basic inference.

Ethical & Responsible Use

This model inherits capabilities and risks from the base Flux family. You are responsible for complying with local laws, platform policies, and the license of the base model. Avoid generating harmful, biased, or misleading content. Apply additional filters if deploying to end users.

License

Usage of these weights inherits the license and usage restrictions of:

the base model (Freepik/flux.1-lite-8B) and
the reward model used during alignment (PickScore).
Review those licenses before commercial use or redistribution.

Acknowledgements

Freepik / Flux team for the original flux.1-lite-8B.
PickScore authors for open-sourcing a practical preference signal.
The broader community for tools in Diffusers & PyTorch that make rapid iteration possible.

FAQ

Q: Why only the transformer?
A: Keeping this repo to the transformer keeps downloads light and makes it easy to swap into existing Flux pipelines.

Q: Do I need to change schedulers or tokenizers?
A: No. Use the same components shipped with the base Flux 1‑Lite 8B pipeline.