ernie / turbo / createImage

Every popular 2026 base shipped a distilled turbo sibling, and ERNIE's is the honest case for the trade: 8 steps, guidance 1.0, dramatically faster - with the quality ceiling lower than the SFT model. Civitai Image Gen (ernie / turbo / createImage) is that variant as a cloud node. If you want ERNIE's layout-and-text strengths at speed, this is the node.

What it is

The turbo text-to-image operation for the ERNIE ecosystem. Structurally it's a mirror of the base ERNIE node with the sampling defaults rewritten for distillation: steps defaults to 8 (vs 20 on the base node) and cfg_scale defaults to 1.0 (vs 4). The KB's model-card facts confirm those defaults match Baidu's own: Turbo is DMD+RL distilled, 8 steps, guidance 1.0 - against the SFT's 50 steps at 4.0.

Required: prompt. Optional: output_format, image_metadata, negative_prompt, sampler/scheduler, seed, quantity (1–12), loras, diffusion_model (CIVITAI_AIR), width/height (64–2048), api_config. Outputs: images, errors, workflow_id, raw_json.

How it fits

The distilled-vs-base question is the same one the KB runs through for every model family: turbo variants converge in a handful of steps and run guidance-off, but the base model has the headroom. For ERNIE specifically, the community found the turbo demos crisp and fast - the "Ernie is absolute masterpiece" showcase was running 8-step Turbo - while the "+222 it's just fine" correction came in on the same thread. The honest read: turbo is the draft-and-iterate pick, and it's the one that makes sense on the cloud, where speed is money.

One trap to avoid: don't crank steps up on this node thinking you're getting base-model quality. The model is distilled for 8 steps; pushing to 50 fights the distillation instead of adding fidelity. If you need the SFT's ceiling, use the base ernie / ernie / createImage node with more steps.

Install & gotchas

Standard pack install - ComfyUI Manager (search Civitai Comfy Nodes), comfy node registry-install civitai-comfy-nodes, or git clone. Auth via env var / OAuth / Civitai Auth node.

And the usual cloud-node notes, briefly: every call is billed in Buzz (quantity 12 = 12 bills), the same layout-and-text strengths and the same model flaws (ethnic bias, grid-pattern artifact) carry over from the base model, and there's no ControlNet socket on the ERNIE surface. The model's own quirks are not the node's fault - but they're not fixed by switching engines, either.

One-line verdict: it's the fast ERNIE, for when you want the layout specialist's output at draft speed. Reaching for it is a statement that you're doing infographics or text-in-image work, not character renders - for those, this pack's other engines are better homes for your Buzz.