Z-Image Turbo Text-to-Image + SDA Diversity LoRA

Z-Image Turbo text-to-image with a LoRA that breaks the "same pose, same angle" problem.

Type a prompt, change the seed, and you get genuinely different compositions instead of variations on one image. The SDA LoRA (zit_sda_v1) rotates the model's flow direction to restore the diversity that distilled few-step models tend to lose.

Defaults are dialed in: 10 steps, CFG 1, 1080 x 1920 portrait. Write your prompt, hit Run.

How do you fix diversity collapse in Z-Image Turbo?

Stack the SDA LoRA on top of Z-Image Turbo at strength 1.0, keep CFG at 1, and run 10 steps with Euler. Now changing the seed changes the image, with different poses, camera angles, and layouts instead of the same composition slightly recolored. Edit your prompt in the CR Prompt Text node.

Prompt (CR Prompt Text node) This is where you type your prompt. It feeds into the text encoder automatically. Z-Image Turbo handles both English and Chinese well, including rendering text inside images like signs, posters, and captions.

SDA LoRA strength (default 1.0) Full strength gives you the most diversity across seeds. Drop to 0.5 or 0.7 if you stack other Z-Image Turbo LoRAs and see structural artifacts or blurring. The two LoRAs can fight over the model's flow direction.

Steps (default 10) Z-Image Turbo is distilled, so 8 to 10 steps is the sweet spot. Going higher burns time without much quality gain. Going under 8 starts to break the output.

CFG (default 1) Z-Image Turbo is CFG-free, so 1 is correct. The negative prompt is zeroed out for the same reason. Do not push CFG higher unless you swap to the base Z-Image model.

Seed (easy seed node) This is the variable that matters now. With SDA on, two different seeds give you two genuinely different images. Set it to "increment" and queue a batch to explore options fast.

Resolution (1080 x 1920 default) Portrait 9:16, sized for social formats. Switch to 1920 x 1080 for landscape or 1024 x 1024 for square. Z-Image Turbo handles non-standard resolutions cleanly.

What is Z-Image Turbo with SDA good for?

Use this when you need visual variety from a fast model. Z-Image Turbo runs in sub-second territory on capable GPUs. The SDA LoRA gives you genuine seed variation instead of near-duplicates. Good for ideation, mood boards, batch generation, and any time you need to pick between options.

Concept artists batching for variety, content creators generating social posts where each frame should feel distinct, and anyone running a prompt across ten seeds to find the best composition. With base Z-Image Turbo, those ten seeds often look like variants of one image. With SDA on, they look like ten different shots.

The tradeoff: more anatomy errors. Distilled models hold body structure together by staying close to a learned trajectory. SDA pushes off that trajectory to get diversity, so hands, limbs, and poses go wrong more often. Plan to filter outputs or use a face detailer pass downstream.

When to skip it: if you need consistent character work, or if you already get the composition you want at seed zero, base Z-Image Turbo is faster and cleaner. SDA is for when sameness is the problem you are trying to solve.

FAQ

What is the SDA LoRA in Z-Image Turbo? SDA is a LoKr adapter that rotates Z-Image Turbo's velocity field to restore diversity across seeds. Distilled few-step models tend to collapse onto similar compositions regardless of seed. SDA pushes the prediction direction off the rectified trajectory, so seed changes produce genuinely different images instead of recoloured duplicates.

What strength should I use for the Z-Image Turbo SDA LoRA? 1.0 if you are running SDA alone, which is this workflow's default. Drop to 0.5 to 0.7 if you stack other Z-Image Turbo LoRAs and see structural artifacts or blurring. The directional gradients can clash. Lower SDA weight means less diversity but cleaner output.

Why is the negative prompt empty in this Z-Image Turbo workflow? Z-Image Turbo runs at CFG 1, which means no classifier-free guidance. Negative prompts only work when CFG is greater than 1. The ConditioningZeroOut node here passes an empty negative to the sampler as the architecture requires. Do not change this unless you swap to the base Z-Image model.

Why does Z-Image Turbo with SDA produce weird hands or limbs sometimes? It is the trade-off SDA makes. To get diversity, it pushes the model off its trained trajectory. Distilled models hold anatomy together by staying close to that trajectory, so going off it produces more limb and hand errors. Lower the LoRA strength to 0.7 to soften this.