ernie / ernie / createImage

ERNIE-Image (Baidu's 8B, Apache 2.0) had the most honest reception arc of any 2026 model: a hype spike, a "+222 it's just fine" correction, and then quiet - while the community kept using Z-Image and Klein. The thing that survived the cycle is a real competence: structured layout and text-in-image. This node runs it on Civitai's cloud, and it's worth having specifically for the jobs ERNIE actually wins.

What it is

The text-to-image node for the ERNIE ecosystem. Required input: prompt. Optional: output_format, image_metadata, negative_prompt, sampler/scheduler, steps (default 20), cfg_scale (default 4), seed, quantity (1–12), loras, diffusion_model (an optional CIVITAI_AIR socket - note the field name here, diffusion_model, differs from anima's diffuser_model), width/height (64–2048), api_config.

Outputs: images, errors, workflow_id, raw_json.

Where it genuinely shines

The KB's ERNIE essay is blunt about the split: image quality is fine, prompt following is not great, but structured layout and text-in-image are the model's pocket. Multi-panel comics, posters, infographics, labelled diagrams - that's the case for ERNIE. "If you are generating a single portrait, there is no case for ERNIE." So the mental model: reach for this node when you're building something with text and layout in the frame, not when you want a pretty character render.

Two more honest warnings from the KB's research: ERNIE shipped with an ethnic bias strong enough that a community LoRA exists purely to produce non-Asian faces, and a grid-pattern noise artifact shows up in most generations. Neither is fixed by running it on the cloud - they're the model's, not the node's.

Install & gotchas

Standard pack install - ComfyUI Manager (search Civitai Comfy Nodes), comfy node registry-install civitai-comfy-nodes, or git clone + pip install -r requirements.txt. Auth via CIVITAI_API_TOKEN, OAuth, or a Civitai Auth node.

Things to know:

1. Steps matter here. The KB's model-card facts note the SFT variant wants 50 steps at guidance 4.0; the node's 20-step default is a starting point, not the model's ideal. If output looks soft, raise steps before touching anything else.
2. No ControlNet socket. The ERNIE surface exposes loras but not control_nets - the cloud engine doesn't take guided structure for this model. Plan accordingly.
3. Metered, per image. quantity up to 12 = 12 billed calls; raw_json shows the cost.
4. The prompt enhancer caveat applies if you ever route prompts through an LLM: Baidu's own writeup says the enhancer trades attribute-binding precision for elaborative detail. If your prompt is already specific, keep it direct.

One take: this node is a specialist's tool. Most people will get more from the pack's Flux 2 or Anima nodes day-to-day. But the moment you need a labelled infographic or a multi-panel layout with legible text, ERNIE's the one model in this pack that was actually built for it - and the cloud version means you don't have to install Baidu's stack to try.