Steppe-Wan2.1: latent-space upscaler for Wan 2.1/Krea 2/Qwen Image, found by agent search

TensorForger · reddit · 2026-11-22

TensorForger released Steppe-Wan2.1, an open-source generative latent upscaler for Wan 2.1, Krea 2 and Qwen Image latents. Unlike two-stage upscale-then-denoise pipelines, its output decodes directly — faster and higher fidelity. It handles images and videos (spatial + temporal upscaling), offers sharpness control via High-Frequency Features conditioning without prompts, and ships in three sizes (200M multi-step, 200M distilled single-step 45ms per 2x, 20M light for real-time). Architecture is local in space and time (chunked window attention, causal temporal path) enabling streaming. Notably, the architecture was discovered by an agentic autoresearch loop — Claude Opus 5.5 running 120 experiments across 8 GPUs to maximize patchfid under a 30-min RTX5090 budget. ComfyUI nodes, HF weights and a demo Space are available.

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

Original post →

More from Multimodal

Multimodal channel →