Descriptions:
Nerdy Rodent puts Baidu’s ERNIE Image Turbo model through a structured head-to-head comparison against Z-Image Turbo in ComfyUI, using a dual-loader workflow that allows toggling between models under identical generation conditions. The tutorial is built around the channel’s “rodent method” of color-coded workflow groupings, making it easy to follow and adapt.
Key technical findings: ERNIE Image Turbo requires 8 steps at CFG 1, produces decent output at lower resolutions, but shows rougher skin and fine-detail texture under close inspection compared to Z-Image — a quality gap that doesn’t improve with increased sampling steps, suggesting it’s an inherent model limitation rather than a tuning issue. For image-to-image work, the denoise threshold is substantially higher than most diffusion models, requiring values above 0.85–0.90 for meaningful transformations. The video also covers ERNIE’s built-in 3B-parameter prompt enhancer, noting it improves output quality across other models like Z-Image as well — and flagging that Baidu’s official benchmark comparisons only used the prompt enhancer with ERNIE, skewing results in its favor.
Beyond text-to-image, the tutorial walks through inpainting, latent upscaling, and image upscaling with practical scheduler guidance (linear quadratic over simple for high-resolution latent upscales). Recommended starting resolution is 1536×1536, with upscaling to 1920×1920 shown to perform well. The full ComfyUI workflow is available via the channel’s Patreon, with model weights on the official ComfyUI Hugging Face repository.
📺 Source: Nerdy Rodent · Published April 17, 2026
🏷️ Format: Tutorial Demo
