Real Dream Flux V1 V4 (RDV4)

A fine-tuned FLUX.1-dev transformer model optimized for high-quality image generation with improved realism and detail.

Model Description

Real Dream Flux V1 V4 is a specialized version of FLUX.1-dev trained to produce photorealistic and detailed imagery. This model contains the transformer component in safetensors format, making it compatible with various training and inference tools including Ostris AI Toolkit.

Model Details

• Model Type: FLUX Transformer (FluxTransformer2DModel)
• Base Model: black-forest-labs/FLUX.1-dev
• Format: safetensors (15.91 GB)
• Architecture:
  • 19 dual attention layers
  • 38 single attention layers
  • 24 attention heads with 128 dims per head
  • Joint attention dimension: 4096

Usage

With Diffusers

from diffusers import FluxPipeline, FluxTransformer2DModel
import torch

# Load the custom transformer
transformer = FluxTransformer2DModel.from_pretrained(
    "SD2704/RDV4",
    subfolder="",
    torch_dtype=torch.bfloat16
)

# Create pipeline with custom transformer
pipe = FluxPipeline.from_pretrained(
    "black-forest-labs/FLUX.1-dev",
    transformer=transformer,
    torch_dtype=torch.bfloat16
)
pipe.to("cuda")

prompt = "your prompt here"
image = pipe(prompt, num_inference_steps=50, guidance_scale=3.5).images[0]
image.save("output.png")

For Training with Ostris AI Toolkit

This model is fully compatible with Ostris AI Toolkit GUI for fine-tuning and LoRA training:

1. In Ostris AI Toolkit, select "Load Model from Hugging Face"
2. Enter model ID: SD2704/RDV4
3. The toolkit will automatically download the transformer and config
4. Configure your training parameters and dataset
5. Start training!

Training Configuration

Recommended settings for fine-tuning:

• Learning rate: 1e-5 to 5e-6
• Batch size: 1-2 (depending on VRAM)
• Gradient accumulation: 4-8 steps
• Mixed precision: bf16
• Optimizer: AdamW or AdamW8bit

Training Data

This model was trained on curated datasets focusing on photorealistic imagery and detailed compositions.

Intended Use

• High-quality image generation
• Fine-tuning for specialized domains
• LoRA training for style adaptation
• Research and creative applications

Limitations

• Requires significant VRAM for inference (minimum 16GB recommended)
• Based on FLUX.1-dev architecture limitations
• May require prompt engineering for optimal results

License

This model follows the FLUX.1-dev licensing terms. Please refer to the base model license for usage restrictions.

Citation

@misc{rdv4,
  author = {SD2704},
  title = {Real Dream Flux V1 V4},
  year = {2024},
  publisher = {HuggingFace},
  url = {https://huggingface.co/SD2704/RDV4}
}

Acknowledgments

Based on FLUX.1-dev by Black Forest Labs.