blanchon / dc_flux_krea_diffusers

blanchon/dc_flux_krea_diffusers

Diffusers-compatible port of DC-Gen-FLUX (Krea) for efficient high-resolution text-to-image generation (2K / 4K).

This repository repackages the original DC-Gen FLUX.1-Krea checkpoint into a 🧨 Diffusers DiffusionPipeline , enabling standard Diffusers workflows while preserving the behavior and performance of the upstream model.

---

Model Details

Model Description

FLUX.1 DC-Gen Krea [dev] is a DC-Gen–adapted FLUX.1-Krea checkpoint that replaces the original FLUX VAE with a deeply compressed DC-AE latent space .
Using embedding alignment followed by lightweight LoRA fine-tuning , DC-Gen enables much faster native 2K / 4K image generation while preserving the base model’s realism and text-rendering quality.

This repository does not retrain the model. It only provides a Diffusers port of the upstream checkpoint for easier inference and deployment.

• DC-Gen method & model: NVIDIA DC-Gen team
  (Wenkun He*, Yuchao Gu*, Junyu Chen*, Dongyun Zou, Yujun Lin, Zhekai Zhang, Haocheng Xi, Muyang Li, Ligeng Zhu, Jincheng Yu, Junsong Chen, Enze Xie, Song Han, Han Cai)
• Diffusers port: @blanchon
• Model type: Text-to-image diffusion (FLUX family, rectified flow transformer)
• License: FLUX.1 [dev] Non-Commercial License (same as upstream)
• Upstream checkpoint: dc-ai/dc_flux_2K4K
• Base model family: black-forest-labs/FLUX.1-Krea-dev

---

Model Sources

• DC-Gen project: https://github.com/dc-ai-projects/DC-Gen
• DC-Gen homepage: https://hanlab.mit.edu/projects/dc-gen
• Paper: https://arxiv.org/abs/2509.25180
• Upstream checkpoint: https://huggingface.co/dc-ai/dc_flux_2K4K
• FLUX.1-Krea base model: https://huggingface.co/black-forest-labs/FLUX.1-Krea-dev

---

Uses

Direct Use

• High-resolution text-to-image generation (1024 → 4096 px)
• Diffusers-based inference, demos, and deployment
• Research on efficient latent-space diffusion and high-resolution synthesis

Downstream Use

• Further research or finetuning only if compliant with the upstream license
• Integration into non-commercial creative or research tools

Out-of-Scope Use

• Commercial usage (not permitted by the FLUX.1-dev license)
• Illegal, harmful, or deceptive content generation

---

Bias, Risks, and Limitations

• The model may reproduce societal biases present in its training data.
• High-resolution generation is GPU- and VRAM-intensive.
• Outputs are not guaranteed to be factual or safe without moderation.
• This repo does not introduce new safety mechanisms beyond those of the base model.

Recommendations

• Review the FLUX.1-dev non-commercial license carefully before use.
• Apply standard content filtering and safety practices in downstream applications.
• Expect memory usage to scale significantly with resolution.

---

How to Get Started with the Model

Minimal Load

import torch
from diffusers import DiffusionPipeline

pipe = DiffusionPipeline.from_pretrained(
    "blanchon/dc_flux_krea_diffusers",
    trust_remote_code=True,
    torch_dtype=torch.bfloat16,
).to("cuda")

Image Generation Example

import torch
from diffusers import DiffusionPipeline

pipe = DiffusionPipeline.from_pretrained(
    "blanchon/dc_flux_krea_diffusers",
    trust_remote_code=True,
    torch_dtype=torch.bfloat16,
).to("cuda")

prompt = "a tiny astronaut hatching from an egg on mars"

image = pipe(
    prompt=prompt,
    width=2048,
    height=2048,
    guidance_scale=4.5,
    num_inference_steps=28,
    output_type="pil",
).images[0]

image.save("dc_flux_krea.png")

For reproducible results, pass a seeded torch.Generator(device="cuda") .

---

Training Details

Training Data

This repository does not introduce new training data.

According to the DC-Gen paper, post-training uses synthetic data generated from the base model to adapt it to a deeply compressed latent space.

Training Procedure

DC-Gen applies:

1. Embedding alignment to bridge the representation gap between latent spaces
2. LoRA fine-tuning to recover base-model quality

See the DC-Gen paper for full methodological details.

---

Evaluation

This repository does not add new evaluation results.

All reported quality, throughput, and latency benchmarks originate from the DC-Gen technical report.

---

Technical Specifications

Architecture

• FLUX-family text-to-image diffusion model
• Rectified flow transformer
• Deeply compressed DC-AE latent space (DC-Gen)

Hardware Requirements

• CUDA-capable GPU strongly recommended
• 2K/4K generation requires substantial VRAM (≥24 GB recommended)

---

Citation

If you use this model in research, please cite:

@article{he2025dc,
  title={DC-Gen: Post-Training Diffusion Acceleration with Deeply Compressed Latent Space},
  author={He, Wenkun and Gu, Yuchao and Chen, Junyu and Zou, Dongyun and Lin, Yujun and Zhang, Zhekai and Xi, Haocheng and Li, Muyang and Zhu, Ligeng and Yu, Jincheng and others},
  journal={arXiv preprint arXiv:2509.25180},
  year={2025}
}

---

Model Card Authors

• DC-Gen research & model: DC-Gen team (NVIDIA)
• Diffusers port & model card: @blanchon

Model Card Contact

• For research questions: see the DC-Gen project page
• For Diffusers port issues: use the Hugging Face Discussions tab