Z-Image-Turbo ControlNet released — 6-block ControlNet, inpaint mode, Canny/HED/Depth/Pose/MLSD (ModelScope)

TL;DR

ModelScope (Alibaba) released a new ControlNet variant for Z-Image-Turbo — the 8-NFE distilled image foundation model from Tongyi-MAI’s Z-Image: An Efficient Image Generation Foundation Model with Single-Stream Diffusion Transformer. The release adds 6-block ControlNet trained from scratch on 1M images at 1328-px resolution, with five condition modalities (Canny / HED / Depth / Pose / MLSD) plus an inpaint mode. Recommended control_context_scale of 0.65–0.80; model, demo, and code shipped together on ModelScope.

Key claims

• 6-block ControlNet variant (vs. typical full-depth ControlNet copies) trained from scratch on 1M images [tweet].
• Supports five structural conditioning modalities — Canny / HED / Depth / Pose / MLSD — and a separate inpaint mode [tweet].
• Recommended control_context_scale range 0.65–0.80 [tweet].
• Resolution: 1328-px training (the same edge length Z-Image-Turbo and similar contemporary 1k-class image models target) [tweet].
• Built on top of Z-Image-Turbo (the 8-NFE Decoupled-DMD + DMDR distilled variant of Z-Image: An Efficient Image Generation Foundation Model with Single-Stream Diffusion Transformer); release ships model + demo + code on ModelScope [tweet].

Method

A community / Alibaba ControlNet adapter for the Z-Image-Turbo base. The novel detail relative to standard ControlNet is the “6-block” trunk — fewer trainable blocks than the original Stable-Diffusion ControlNet’s full-depth trainable-copy recipe, which lowers memory and probably suits the few-step distilled Turbo backbone. Conditioning channels follow the standard structural-control family: edges (Canny / HED / MLSD), depth, and pose. An “inpaint mode” is bundled alongside the structural conditions — likely a separate channel/head rather than a sixth structural map. No paper or technical report accompanies the announcement; the tweet links to ModelScope and a GitHub repo under aigc-apps/Vide… (truncated in the tweet preview).

Results

None reported in the tweet — this is a release announcement, not an evaluation. The single concrete tuning knob shared is control_context_scale ∈ [0.65, 0.80].

Why it’s interesting

Two reasons to track this for Luma. First, it’s the first public structural-control adapter for Z-Image, which Z-Image: An Efficient Image Generation Foundation Model with Single-Stream Diffusion Transformer argued was the new cost-quality frontier in open text-to-image at 6B parameters; pose / depth / Canny adapters are the practical interfaces creative pipelines actually wire into. Second, the “6-block ControlNet” trained against an 8-NFE distilled backbone is a useful datapoint on how much trainable depth you actually need on top of a few-step student — a question that also matters for adapter design on the diffusion-distillation cluster (Diffusion Distillation, pi-Flow: Policy-Based Few-Step Generation via Imitation Distillation, Adversarial Flow Models). Compare the architecture-level multimodal-conditioning approach in C3Net: Compound Conditioned ControlNet for Multimodal Content Generation, which tries to unify many control signals inside one trainable copy rather than ship multiple per-condition adapters.

See also

• Z-Image: An Efficient Image Generation Foundation Model with Single-Stream Diffusion Transformer — the base model this ControlNet attaches to; explains the S3-DiT architecture and the Z-Image-Turbo 8-NFE distillation pipeline.
• C3Net: Compound Conditioned ControlNet for Multimodal Content Generation — earlier multi-modal ControlNet design; useful contrast for “one adapter, many conditions” vs. ModelScope’s “many adapters, many conditions.”
• Open foundation-model releases — Z-Image-Turbo ControlNet fits the same open-release-with-code+weights+demo pattern as the underlying Z-Image release.
• Diffusion Distillation — Z-Image-Turbo is the 8-NFE distilled student; structural conditioning on top of few-step distilled backbones is an open design question.