Descriptions:
Veteran AI takes a systematic look at Z-Image Turbo ControlNet 2.0, released by Alibaba’s PI team, and isolates the most practically significant improvement over version 1.0: the existence of a reliable intermediate control state. By running multiple generations at ControlNet weights of 1.0, 0.75, and 0.5 with both versions side by side, the presenter demonstrates that 1.0 produces wildly inconsistent poses once the weight drops below 1.0 — effectively behaving as an on/off switch — while 2.0 maintains coherent pose fidelity down to 0.5 weight and degrades gracefully below that, giving users meaningful control over the strength of pose guidance.
The workflow setup is consistent throughout: Z-Image Turbo main model, DW OpenPose for pose extraction, a custom character LoRA, CFG 3.0, and an 8-step dual sampler (4 steps for composition, 4 for refinement). This controlled setup makes the behavioral difference between versions clearly attributable to the ControlNet model itself rather than other variables.
The video also covers an important caveat: the Pose + Inpaint feature advertised as the headline improvement in ControlNet 2.0 is not yet implemented in official ComfyUI nodes. The presenter backs this up with a specific GitHub commit reference showing developers have explicitly deferred inpaint support, saving viewers time chasing a feature that doesn’t yet work. Practical setup notes — including the non-obvious model_patches directory for placing the ControlNet 2.0 file — make this useful for anyone trying to reproduce the results immediately.
📺 Source: Veteran AI · Published December 16, 2025
🏷️ Format: Benchmark Test
