Steppe-Wan2.1: open-source latent upscaler for Wan/Qwen/Krea, arch found by 120 Claude-run experiments

TensorForger · reddit · 2026-11-22

TensorForger released Steppe-Wan2.1, a generative latent upscaler for Wan 2.1, Krea 2, and Qwen Image latent spaces — weights, ComfyUI nodes, and an HF demo space are all open.

Key idea: instead of the classic two-stage "deterministic upscale → denoise with the base model," Steppe upscales 2x directly in latent space and decodes immediately — faster, with better fidelity and preservation of low-res information.

Features:

Notable: the architecture was discovered by an agentic autoresearch loop inspired by karpathy/autoresearch — Claude Opus 5.5 ran 120 experiments in parallel on 8 GPUs (30-min RTX5090 budget per run, then scaling the winners), landing on a Swin-transformer + conv hybrid with novel training-stability tricks. Experiment logs are in the dev repo.

Related event: Steppe-Wan2.1 Open-Sources One-Step Latent-Space Upscaling(2 posts)→

Original post →

More from Multimodal

Multimodal channel →