Descriptions:
Nerdy Rodent explores Z-Image Base inside ComfyUI, the slower but more capable sibling of Z-Image Turbo, focusing on what the Base model unlocks that Turbo cannot: negative prompting through a CFG guider node and proper LoRA training support. The workflow is adapted directly from the Turbo build, with changes concentrated on swapping in the CFG guider across stage-one samplers, the detailed demon upscaler group, and the high-res fix group — keeping the upgrade path straightforward for existing Turbo users.
The tutorial tests multiple CLIP model variants (Qwen 34B FP8 mixed and FP4 mixed), shift value adjustments, split sigma configurations, and stage-two upscaling with the high-res fix. Side-by-side comparisons via ComfyUI’s compare node make the differences between configurations easy to evaluate. Older Z-Image Turbo LoRAs are shown to partially transfer to the Base model, with split sigma sampling producing noticeably better compatibility than the standard K-sampler path — a practical data point for users with existing LoRA libraries.
A critical practical note: enabling Sage Attention when running Z-Image Base causes visible image artifacts, and the presenter flags this clearly before the walkthrough begins. The video runs at CFG4 and 25 steps for approximately 45 seconds per image, making it a useful reference for anyone weighing the speed-quality tradeoff between the two Z-Image variants.
📺 Source: Nerdy Rodent · Published January 28, 2026
🏷️ Format: Tutorial Demo
