AI Models

Machine learning models I've trained, fine-tuned, quantized, or otherwise worked on.

• Krea 2 Depth ControlNet Author
  Task: Depth-to-Image Base: krea/Krea-2 Repo: Patil/Krea-2-depth-controlnet

  A depth-conditioned ControlNet for the Krea 2 diffusion model, letting you steer text-to-image generation with a depth map so the output respects the spatial layout and 3D structure of a reference. Feed in a depth map (from an existing image or a depth estimator) alongside your prompt and the ControlNet keeps foreground/background relationships and object placement intact while the base model handles style and content, making it well suited for consistent scene composition, relighting, and structure-preserving image variations.

  • Diffusion
  • ControlNet
  • Depth
  • Text-to-Image
  • Structure Control
• Krea Turbo SVDQuant Author
  Task: Text-to-Image Base: krea/Krea-2-Turbo Repo: Patil/krea-turbo-svdquant

  A 4-bit (W4A4) SVDQuant build of the Krea Turbo diffusion transformer, packaged for the Nunchaku runtime. SVDQuant pushes both weights and activations down to 4-bit while absorbing the resulting outliers with a low-rank branch, so the model keeps near-FP16 image quality at roughly a quarter of the memory and a sizeable inference speed-up. Turbo distillation on top means high-fidelity text-to-image generation in just a handful of sampling steps, making it practical to run on consumer GPUs.

  • Diffusion
  • SVDQuant
  • 4-bit
  • Quantization
  • Text-to-Image
• Tiny Klein 3 VAE Author
  Task: Latent Decoding Base: madebyollin/taef2 Repo: Patil/tiny-flux-klien-3-vae

  A 2.7M-parameter tiny decoder for FLUX 3 Image Klein latents, fine-tuned from TAEF2. FLUX 3 Klein ships the FLUX.2 autoencoder byte-for-byte, so the tiny model reads the same latents; distillation on 25.8k real photos with L1, LPIPS, a light PatchGAN loss and EMA weights, first at 512px and then 1024px, has a 26–32% smaller perceptual (LPIPS) gap to the full 84M-parameter VAE than TAEF2 on FLUX 3 generations. It decodes about 11x faster with roughly 4x less memory, for live previews, high-resolution decoding and memory-limited GPUs.

  Read the write-up →
  • VAE
  • Distillation
  • FLUX 3
  • Tiny Autoencoder
  • Fast Decoding