z-image-turbo-sda — Hugging Face model release¶

Development line: project:z-image-turbo-sda · thread hugging-face-model-release
Last event: 2026-03-26 · 1 dated since 2026-03-26 · Researched: 2026-09-06 · confidence: medium

What it is¶

Z-Image-Turbo-SDA is a third-party LoKr/LyCORIS adapter for people running Tongyi-MAI/Z-Image-Turbo, not a standalone Z-Image checkpoint.

• Loads into the Turbo pipeline to vary pose, camera angle, and layout between seeds.
• Retains an 8-step workflow, with a reported diversity-versus-anatomy trade-off.

Development line¶

• 2026-03-26 — z-image-turbo-sda was linked to a Hugging Face model repository. Z-Image-Turbo-SDA was referenced as an already-published adapter. No repository commit is recorded on this date.

What changed¶

• 2026-03-11: The repository uploaded zit_sda_v1.safetensors and documented a LoKr adapter for Tongyi-MAI/Z-Image-Turbo.
• 2026-03-26: Z-Image-Turbo-SDA was referenced as an already-published adapter; no repository commit is recorded on this date.
• 2026-04-09: Two commits labelled Update README.md changed documentation, not the public weight artifact.
• 2026-05-12: The maintainer acknowledged compatibility issues with some derivative ZIT models and pointed to a separate code-only alternative.

How to use this¶

From 2026-03-26, we treat the Hugging Face repository F16/z-image-turbo-sda as the distribution reference for z-image-turbo-sda. Check its model card, files, license, and compatibility before use.

1. Start with the exact Tongyi-MAI/Z-Image-Turbo base. SDA is an adapter, not a checkpoint to load by itself. — https://huggingface.co/F16/z-image-turbo-sda
2. In a CUDA Diffusers setup, load the base in float16, then load F16/z-image-turbo-sda with pipeline.load_lora_weights(..., adapter_name="sda_diversity"). The card calls for current diffusers plus peft. — https://huggingface.co/F16/z-image-turbo-sda
3. First reproduce the supplied reference run: 8 inference steps, CFG 1.0, and a fixed grid of seeds. Assess variation across the grid rather than from one output. — https://huggingface.co/F16/z-image-turbo-sda

Best practices¶

• Validate SDA alone before stacking other Z-Image-Turbo LoRAs. If stacking is needed, the maintainer suggests reducing SDA weight to 0.5–0.7 and accepting less diversity. — https://huggingface.co/F16/z-image-turbo-sda
• Use a fixed multi-seed A/B test with simple and spatially constrained prompts, then inspect complex poses for anatomy errors. The card notes that constrained prompts leave less room for composition changes. — https://huggingface.co/F16/z-image-turbo-sda
• Keep the adapter card's 8-step, CFG-1.0 preset separate from the official base recipe. The official recipe specifies ZImagePipeline, bf16, 9 scheduler steps for 8 DiT forwards, and CFG 0.0. — https://huggingface.co/Tongyi-MAI/Z-Image-Turbo
• Treat the reported LPIPS improvement and diversity-recovery percentages as maintainer results, not independently reproduced benchmarks. — https://huggingface.co/F16/z-image-turbo-sda

Superseded by this¶

• Nothing marked obsolete yet.

Still unknown¶

• No public training code, dataset provenance, immutable training-checkpoint digest, minimum diffusers/peft version, or supported ComfyUI workflow was found.
• The model card's LPIPS values and claimed 70.2%/58% diversity recovery are maintainer-reported and were not independently validated.
• No public repository activity establishes a new SDA release exactly on 2026-03-26. The date identifies an existing artifact rather than a demonstrated weight update.
• No evidence was found that Tongyi-MAI created, endorsed, or maintains this adapter.

Sources¶

Agent brief¶

• Subject: project:z-image-turbo-sda, thread hugging-face-model-release, 1 dated events 2026-03-26 → 2026-03-26.
• Practical note: From 2026-03-26, we treat the Hugging Face repository F16/z-image-turbo-sda as the distribution reference for z-image-turbo-sda. Check its model card, files, license, and compatibility before use.
• Confidence: medium. Dated supersedes above are the authority for what is obsolete.