PixelWave Flux.1-dev 03 is a general-purpose diffusion model fine-tuned on over 5000 diverse images for 5+ weeks using Kohya by creator Mikey And Friends. The model consistently produces outputs that closely match the prompt descriptions. It is able to generate images in various artistic styles with superior photorealism compared to the original Flux.1-dev model.
Table of Contents
Creating PixelWave Flux.1-dev 03
Mikey trained the FLUX.1-dev 03 model using his powerful RTX 4090 GPU and custom configuration files for the Kohya fine-tuning library. Starting from the FLUX.1-dev checkpoint, he experimented with different hyperparameters like learning rates and gradient accumulation steps. Mikey also developed a strategy of training in batches to test outputs after each checkpoint and prevent overfitting.
Fine-tuning of Flux.1-dev
Key settings for PixelWave Flux.1-dev 03’s 5+ week training include:
- Learning rate of 1.8e-6
- Adafactor optimizer
- BF16 and FP8 mixed precision
- Cache latent vectors for faster training
- Multi-resolution sampling from 256 to 2096p
- Up to 1000 steps per image
- Save checkpoints every 1200 iterations
Image Dataset for PixelWave Flux.1-dev 03
To maximize diversity, Mikey selected over 5000 photos without AI generation. Genres included nature, architecture, portraits and more. Kohya’s bucketing system automatically resized images to different aspect ratios during training. This helps the model generalize across resolutions like square, 16:9, 2:3, etc.
Model Performance
After weeks of training, the model learned a wide range of visual styles while maintaining high fidelity. Reviews praise its ability to reproduce artistic styles and photographic techniques. Standard prompts use DPM++ scaling with 2M point sets over 15-30 steps. Upscaling inputs beyond 1024p can produce even higher quality 4K imagery.
PixelWave Flux.1-dev 03 vs. Original Flux.1-dev
As you can see in the examples below, PixelWave Flux.1-dev 03 follows the prompts correctly and is able to generate images in various artistic styles with far greater photorealism and attention to detail than the original Flux.1-dev model. The original Flux.1-dev struggles at times to accurately interpret certain styles, and its images appear comparatively blurrier and less grounded in photographic consistency.
Images Source: imgur
How to Get Started With PixelWave Flux.1-dev 03
PixelWave Flux.1-dev 03 is available on HuggingFace and Civitai with GGUF, FP8 and BF16 formats for different hardware. Recommended settings help match prompts to desired content genres. Additional guides from Mikey cover advanced techniques like skip stepping and latent space manipulations. The creator also developed helpful ComfyUI nodes for GGUF loading and processing.
Resources from Creator
In a Reddit post, Mikey shared valuable insights into training strategies and supported diffusion models and conversion tools for GGUF formats. This open communication helps the community optimize models while respecting non-commercial licenses. Overall, this new fine-tuned model demonstrates how focused input from experienced creators leads to broadly capable AI systems.
| Latest From Us
- Hugging Face CEO Shares His 2025 AI Predictions
- Stanford 2024 AI Index Report Confirms That AI Leaves Human Capabilities in the Dust Across Domains
- World Labs Introduces Spatial AI Model That Lets You Navigate 3D Worlds from 2D Images
- Nous Research Develops DisTrO Powered by Distributed Machines Across the Internet
- Tencent Introduces HunyuanVideo, An Open-Source Triumph in Video Generation Excellence
