blanchon/dc_flux_krea_diffusers

1---2library_name: diffusers3pipeline_tag: text-to-image4tags:5  - text-to-image6  - image-generation7  - flux8  - dc-gen9  - diffusers10base_model:11  - dc-ai/dc_flux_2K4K12  - black-forest-labs/FLUX.1-Krea-dev13---14 15# blanchon/dc_flux_krea_diffusers16 17**Diffusers-compatible port of DC-Gen-FLUX (Krea)** for efficient high-resolution text-to-image generation (2K / 4K).18 19This 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.20 21---22 23## Model Details24 25### Model Description26 27**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**.  28Using **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.29 30This repository does **not** retrain the model. It only provides a **Diffusers port** of the upstream checkpoint for easier inference and deployment.31 32- **DC-Gen method & model:** NVIDIA DC-Gen team  33  (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)34- **Diffusers port:** @blanchon35- **Model type:** Text-to-image diffusion (FLUX family, rectified flow transformer)36- **License:** FLUX.1 [dev] **Non-Commercial License** (same as upstream)37- **Upstream checkpoint:** `dc-ai/dc_flux_2K4K`38- **Base model family:** `black-forest-labs/FLUX.1-Krea-dev`39 40---41 42## Model Sources43 44- **DC-Gen project:** https://github.com/dc-ai-projects/DC-Gen  45- **DC-Gen homepage:** https://hanlab.mit.edu/projects/dc-gen  46- **Paper:** https://arxiv.org/abs/2509.25180  47- **Upstream checkpoint:** https://huggingface.co/dc-ai/dc_flux_2K4K  48- **FLUX.1-Krea base model:** https://huggingface.co/black-forest-labs/FLUX.1-Krea-dev  49 50---51 52## Uses53 54### Direct Use55 56- High-resolution text-to-image generation (1024 → 4096 px)57- Diffusers-based inference, demos, and deployment58- Research on efficient latent-space diffusion and high-resolution synthesis59 60### Downstream Use61 62- Further research or finetuning **only if compliant with the upstream license**63- Integration into non-commercial creative or research tools64 65### Out-of-Scope Use66 67- Commercial usage (not permitted by the FLUX.1-dev license)68- Illegal, harmful, or deceptive content generation69 70---71 72## Bias, Risks, and Limitations73 74- The model may reproduce societal biases present in its training data.75- High-resolution generation is GPU- and VRAM-intensive.76- Outputs are not guaranteed to be factual or safe without moderation.77- This repo does not introduce new safety mechanisms beyond those of the base model.78 79### Recommendations80 81- Review the FLUX.1-dev non-commercial license carefully before use.82- Apply standard content filtering and safety practices in downstream applications.83- Expect memory usage to scale significantly with resolution.84 85---86 87## How to Get Started with the Model88 89### Minimal Load90 91```python92import torch93from diffusers import DiffusionPipeline94 95pipe = DiffusionPipeline.from_pretrained(96    "blanchon/dc_flux_krea_diffusers",97    trust_remote_code=True,98    torch_dtype=torch.bfloat16,99).to("cuda")100````101 102### Image Generation Example103 104```python105import torch106from diffusers import DiffusionPipeline107 108pipe = DiffusionPipeline.from_pretrained(109    "blanchon/dc_flux_krea_diffusers",110    trust_remote_code=True,111    torch_dtype=torch.bfloat16,112).to("cuda")113 114prompt = "a tiny astronaut hatching from an egg on mars"115 116image = pipe(117    prompt=prompt,118    width=2048,119    height=2048,120    guidance_scale=4.5,121    num_inference_steps=28,122    output_type="pil",123).images[0]124 125image.save("dc_flux_krea.png")126```127 128For reproducible results, pass a seeded `torch.Generator(device="cuda")`.129 130---131 132## Training Details133 134### Training Data135 136This repository does **not** introduce new training data.137 138According to the DC-Gen paper, post-training uses **synthetic data generated from the base model** to adapt it to a deeply compressed latent space.139 140### Training Procedure141 142DC-Gen applies:143 1441. **Embedding alignment** to bridge the representation gap between latent spaces1452. **LoRA fine-tuning** to recover base-model quality146 147See the DC-Gen paper for full methodological details.148 149---150 151## Evaluation152 153This repository does not add new evaluation results.154 155All reported quality, throughput, and latency benchmarks originate from the DC-Gen technical report.156 157---158 159## Technical Specifications160 161### Architecture162 163* FLUX-family text-to-image diffusion model164* Rectified flow transformer165* Deeply compressed DC-AE latent space (DC-Gen)166 167### Hardware Requirements168 169* CUDA-capable GPU strongly recommended170* 2K/4K generation requires substantial VRAM (≥24 GB recommended)171 172---173 174## Citation175 176If you use this model in research, please cite:177 178```bibtex179@article{he2025dc,180  title={DC-Gen: Post-Training Diffusion Acceleration with Deeply Compressed Latent Space},181  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},182  journal={arXiv preprint arXiv:2509.25180},183  year={2025}184}185```186 187---188 189## Model Card Authors190 191* **DC-Gen research & model:** DC-Gen team (NVIDIA)192* **Diffusers port & model card:** @blanchon193 194## Model Card Contact195 196* For research questions: see the DC-Gen project page197* For Diffusers port issues: use the Hugging Face Discussions tab198