ERNIE_Turbo Model (CRT AutoDL)

ERNIE-Image is Baidu's Apache 2.0 8B - the model that got a two-week hype spike on real layout-and-text ability, then lost the attention race to Z-Image and Krea 2. But if you want a local model that can actually render structured text and clean infographics, it still has a niche, and CRTAutoDLErnieTurboModel is the zero-effort way to load it: the fp8 Turbo checkpoint downloads itself from Bedovyy's quant repo on first use and comes out as a ready MODEL.

The Turbo variant is the fast, distilled member of the family - the one you'd actually use day-to-day rather than the heavier base ernie-image-fp8. This node loads ernie-image-turbo-fp8.safetensors into ComfyUI/models/diffusion_models/ with the same download-on-first-queue mechanic as every CRT AutoDL loader.

The three inputs

Identical to the other AutoDL model loaders, which is nice if you've met one already:

• sage_attention (default auto) - SageAttention attention when importable; the main speed lever.
• enable_fp16_accumulation (default on) - leave it on.
• patch_cublaslinear (default off) - ComfyUI's --fast cublas patch, off unless you're measuring.

For an 8B model on a consumer card, defaults are genuinely fine. Don't tune before you've seen a baseline.

Wiring

MODEL out feeds a standard sampler. Pair it with CRTAutoDLErnieCLIP (the Ministral-3 3B text encoder) and CRTAutoDLErnieVAE - and note that ERNIE's VAE is the same flux2-vae that Flux 2 uses, which is a nice small consistency win in the AutoDL family. Three nodes, complete stack.

What to know before you commit

• First run downloads ~8GB (fp8). Console progress bar, no UI indicator - be patient on first queue.
• ERNIE's quirks are in the model, not the node. The known weaknesses - an ethnic bias and a grid-pattern artifact - come with the weights. If you hit those, no loader setting fixes them; they're why the model's community shrunk despite the licence.
• There's a second AutoDL ERNIE loader (CRTAutoDLErnieTurboNVFP4Model) for the NVFP4 build if you're on a card that likes it. This one is the plain fp8, which is the safest default.

Install CRT-Nodes via ComfyUI Manager (search CRT-Nodes) or git clone https://github.com/plugcrypt/CRT-Nodes.git into custom_nodes, install requirements.txt, restart, and queue once to trigger the download.