Training Image Models With Adversarial Loss

Ok-Constant8386 · reddit · 2026-07-15

The author shares an experiment training Krea 2 in a GAN-like manner, replacing the MSE loss in diffusion with pure adversarial loss to see if it improves details, textures, and sharpness.

Experimental Setup

Dataset: opendiffusionai/pexels-woman-solo
Base: 1 epoch of SFT on Krea 2 LoRA
Generated 500 fake images, then fine-tuned DINOv3 840M ViT for patch-level real/fake discrimination
Instead of looking at the whole image, the discriminator uses 16×16 patches as classification units
DiffAugment was added during training; the author notes tuning is ongoing and regularization is needed

Training Approach

The generator first generates images using a random timestep schedule
The discriminator judges the authenticity of image patches
The generator tries to fool the discriminator
During discriminator training, real images and images passed through VAE encoding/decoding are used to reduce VAE bias
To balance the adversarial process, the discriminator updates every step, while the generator updates every 3 steps

Author's Take

He believes the MSE objective in diffusion models pushes the model to learn "average pixel values," losing details and textures, which causes blurriness. Introducing adversarial loss could fix this. Current results remain experimental, but the author believes that with large-scale training and more ablations, this route could enhance realism, style learning, and overall quality.

Original post →

Training Image Models With Adversarial Loss

More from Multimodal