Alibaba drops Z-Image base model for diverse generation and easy fine-tuning
Alibaba's Tongyi Lab just open-sourced Z-Image, the undistilled foundation model behind its fast Turbo edition. Built specifically for customization, it fixes the Turbo version's repetitive faces and compositions while adding negative prompt support. It needs 50 steps instead of 28, and ComfyUI already supports it out of the box.
Z-Image is Alibaba Tongyi Lab's newly open-sourced image generation foundation model. It serves as the complete, undistilled base model behind the earlier Z-Image Turbo, engineered specifically for model customization and community fine-tuning.
Fixing Turbo's biggest flaws
Alibaba rolled out Z-Image-Turbo in November 2025. While that model delivered blistering generation speeds and sharp aesthetics, it had serious diversity issues under the hood.
Because Turbo underwent heavy distillation and RLHF, it suffered from repetitive compositions even across different seeds, and group shots frequently rendered identical faces on multiple people. It was basically unusable for downstream model training.
The newly released Z-Image foundation model steps back to the pre-distillation checkpoint. It resolves the repetitive faces and locked-in framing, bringing a much wider range of facial features and styles, especially across illustrative aesthetics.
It also introduces native support for negative prompts, allowing creators to cleanly subtract unwanted elements without hacky workarounds.
The trade-offs and how to run it
More freedom comes with obvious performance costs compared to Turbo.
The required generation steps climb from 28 to 50, slowing down raw generation throughput noticeably.
Out of the box without any LoRA or fine-tuning, its default image quality shifts down from "Very High" to "High." Because it lacks pre-baked aesthetic tuning, getting show-stopping results requires prompt crafting, ControlNet, or custom fine-tunes.
The weights are already live on Hugging Face under the Tongyi-MAI repository. If you rely on node-based workflows, ComfyUI launched native Day 0 support, meaning you can drop the weights straight into your pipeline right away.
Why it matters
If you just want rapid, one-click prompts for casual generation, the faster Turbo edition or a polished commercial API is still your best bet.
However, for LoRA trainers and downstream pipeline builders, this unconstrained base model is what the open-source community actually needed. Distilled models compress latent possibilities for raw speed, but by releasing the uncompromised foundation, Alibaba is giving the open developer ecosystem a reliable canvas to build custom models upon.
Sources (1)
- gigazine.net(opens in a new tab)Originalgigazine.net