--- title: Sample-Aligned Decoders emoji: 🎨 colorFrom: red colorTo: green sdk: gradio sdk_version: 6.32.0 app_file: app.py short_description: Drop-in fine-tuned VAE decoders for latent generators python_version: "3.12" startup_duration_timeout: 30m --- # Sample-Aligned Decoders (SAD) **Paper**: [Sample-Aligned Decoders for Latent Generative Models](https://huggingface.co/papers/2611.22458) · **Code**: [github.com/torpedo2003/SAD](https://github.com/torpedo2003/SAD) · **Decoders**: [huggingface.co/torpedo2003/SAD](https://huggingface.co/torpedo2003/SAD) SAD is a decoder post-training framework for latent generative models: fine-tune **only the VAE decoder** on latents sampled from the frozen generator (~1 GPU-hour). The decoder keeps its architecture, so it is a **drop-in swap with zero extra inference cost** — it corrects color cast, exposure and texture while leaving content and layout untouched. ## What this demo does It runs **PixArt-α 256px** with the authors' exact sampler recipe (20 DPM-Solver++ steps, guidance 4.5 — the configuration the `pixart-alpha-256/sad-gan` decoder was trained for): 1. one prompt → **one latent** from the frozen PixArt-α base model; 2. that same latent is decoded twice — with the **original SD-VAE decoder D₀** and with the **SAD decoder**; 3. the **|SAD − D₀| heatmap** (paper colormap) shows where the decoder changed the image. Because the latent is identical, every difference between the two images is purely the decoder. ## Example prompts The examples are the official PixArt-α prompts from the [PixArt-XL-2-256x256](https://huggingface.co/PixArt-alpha/PixArt-XL-2-256x256) model repo, sampled with the same 20-step / cfg-4.5 recipe used here.