Skip to main content
Back

madebyollin's Texture-Fix VAE Kills Qwen-Image-2.1's Ugly Grid Artifacts

A community fine-tune of the Qwen-Image-2.1 VAE decoder wipes out checkerboard artifacts and boosts perceptual quality with just 7.5M trainable parameters.

  • Drop-in fine-tuned VAE decoder for Qwen-Image-2.1 that removes checkerboard artifacts.
  • rFID on COCO improves from 3.37 to 2.08, with only a small PSNR/LPIPS regression.
  • Trained for ~5000 steps at lr 3e-5, unfreezing only 7.5M parameters in the top decoder stages.
  • Uses the TAESD recipe mixing MSE/MAE, LPIPS, and adversarial losses; base model likely lacked working GAN loss.
  • Works as a one-file swap in ComfyUI or Diffusers; free under Qwen research license.
  • Author also released a distilled tiny AE for realtime latent previews alongside this release.

Independent researcher madebyollin has released Texture-Fix VAE, an unofficial replacement decoder for Qwen-Image-2.1. The fine-tune targets checkerboard patterns and noisy textures in detailed, photo-style images without changing the diffusion backbone.

Because the checkpoint retains Qwen-Image-2.1’s latent format, diffusion-model LoRAs and prompts remain compatible. ComfyUI workflows that expose VAE selection can adopt it through a file swap, while Diffusers users can inject the replacement when constructing the pipeline.

Where the grids come from

A variational autoencoder, or VAE, compresses an image into a smaller latent grid and later reconstructs pixels from that representation. During decoding, learned upsampling expands the grid to full resolution. Periodic errors in that process can appear as checkerboards, especially across foliage, fabric, skin, stone, and other high-frequency textures.

The training objective affects how a decoder handles details that the compressed representation cannot recover precisely. The TAESD recipe used for this fine-tune combines three types of loss:

  • MSE and MAE losses reward pixel-level accuracy but tend to average uncertain details, producing smoother results.
  • LPIPS loss compares learned visual features rather than individual pixels, although it can still tolerate blur or periodic texture errors.
  • Adversarial loss rewards reconstructions that resemble real images, encouraging sharper and more plausible detail. It can also generate detail that differs from the source image.