Flux1.Krea-dev-villeneuve

Prompt
A vast desert landscape with mysterious structures

Prompt
FBI agents breaking to a drug dealer's home

Model description

Overview

Finetuned LORA off FLUX.1-Krea-dev that better captures the style of Denis Villeneuve. The training took ~11 H200 hours to train, and is optimized in a number of ways, including but not limited to: VAE caching, image interpolation, optimized attention via xformers, torch.compile(), Cosine LR annealing. The dataset of ~200 images was curated to especially capture the width of the directorial body of work, with captioning also focusing on style of the model.

Uses

Flux 1 - Krea dev (black-forest-labs/FLUX.1-Krea-dev) as the base model for training
uv for package management
ruff for code quality
ty for type checking
modal for infrastructure
shotdeck (https://shotdeck.com/) for training stills and data
Qwen 2.5VL - 3B for image captioning