ERNIE-Image-Turbo

ERNIE-Image-Turbo is the same story as the full ERNIE-Image, compressed. Where the base model burns 30 steps to get its clean, legible text-in-image, Turbo is a distilled variant that runs in 2–10 steps (default 8). It's the fastest way to draft a layout: posters, cards, mockups with text baked in, quick enough that you can iterate on the composition instead of waiting on it.

As with everything in the Runware pack, "fast" here means "fast on someone else's GPU." You send a prompt, the cloud runs the model, you get an IMAGE tensor back. No model files, no VRAM math.

How it works

Distilled models learn to jump to the finished image in very few denoising steps, and the tradeoff is quality at the margins. Turbo keeps the ERNIE strengths - structured composition and readable text - but you trade away some of the full model's refinement. The node is a thin taskType: imageInference REST call with the model AIR baidu:ernie-image@turbo baked in; all the real logic lives on Runware's side.

Inputs that matter

• positivePrompt - required, and the only thing you must touch. Turbo's speed means you can afford to iterate on it a lot.
• steps - 2–10, default 8. Below 6, text can start smearing; above 8 you're barely gaining anything. 8 is genuinely the sweet spot here.
• settings.promptEnhance.enabled - the LLM prompt expander, off by default. Paired with temperature (1.2) and topP (0.95) if you want it. Worth trying when short prompts come back too literal.
• acceleration - none/low/medium/high. This stacks on top of the distillation, so it's a "how cheap can I make this" knob more than a quality one.
• width / height - up to 2048 in 16px steps.

Note there's no CFGScale on this node - distilled models mostly ignore high guidance anyway, so the pack just doesn't expose it. Output is image → Preview Image or Save Image.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/Runware/ComfyUI-Runware
pip install -r ComfyUI-Runware/requirements.txt

Or ComfyUI Manager → search Runware → Install → restart. API key via ComfyUI Settings → Runware API key or RUNWARE_API_KEY. Prepaid top-up, cost per run shown on the title bar.

Gotchas

• The same caveats as the base ERNIE model: the community flagged an ethnic bias and a grid-pattern artifact. Turbo shows both more aggressively under pressure, so don't push tiny text or dark low-step generations and then wonder what went wrong.
• Fewer steps + JPG output = the fastest way to wreck glyph edges. Keep PNG for anything with text that matters.
• If you outgrow the speed and want the detail back, step up to the full ERNIE-Image node - same pack, same key, same wiring.

The honest pitch: Turbo is the version you use when ERNIE's layout ability matters but the full model's cost and latency don't. Draft fast, polish with the big one.