ERNIE Turbo Model (CRT AutoDL)

ERNIE Turbo Model (CRT AutoDL) is a diffusion-model loader with a specific taste: it pulls Baidu's ERNIE-Image-Turbo checkpoint straight from HuggingFace on first use, drops it into your ComfyUI/models/diffusion_models folder, and hands you a ready-to-sample MODEL. No hunting for a download link, no "put this file in the right folder" dance - you add the node, pick a variant, run, and it fetches.

It's one of the "AutoDL" family in PGCRT's CRT-Nodes pack, and that family is the whole appeal: every model in it is pinned to a known-good URL and auto-downloaded, so a workflow using these nodes just works on a fresh machine as long as you have disk space and patience for the first run.

The ERNIE context, honestly

ERNIE-Image is Baidu's Apache-2.0 image model (8B, single-stream DiT with a fat 128-channel VAE). The Turbo variant is the DMD + RL distilled, 8-step version. The community verdict on ERNIE is worth knowing before you build a workflow around it: it launched in April 2026 to a loud two-week honeymoon and then went quiet while everyone kept using Z-Image and Flux 2 Klein. It's genuinely strong at structured layout, infographics, and text-in-image - and unremarkable at single-portrait photorealism. So this node is the low-friction way to try ERNIE Turbo; whether it earns a permanent slot is a taste call. If you're doing posters or labelled diagrams, it has a real pocket.

What the node actually does

Pick model_name:

• Turbo - the fp8 quant (ernie-image-turbo-fp8.safetensors, from Bedovyy's ERNIE-Image-Quantized repo).
• Turbo NVFP4 - the NVFP4 quant of the same model. NVFP4 is the newer, leaner 4-bit format; if you're on Blackwell and memory-tight, that's the one to try.

The rest of the inputs are the CRT "fast settings" that show up across every AutoDL diffusion loader:

• attention_method - defaults to disabled, which is correct for most people. The menu offers PyTorch attention plus a stack of SageAttention modes (sageattn_qk_int8_pv_fp16_cuda, sageattn3, etc.) if you've installed SageAttention and want to shave VRAM or speed up attention. Don't touch it until you know why.
• patch_cublaslinear - off by default; toggles ComfyUI's cublas_ops fast path for cuBLAS matmuls.
• enable_fp16_accumulation - on by default, matching what the community settled on for this model family.

Output is a single MODEL, which wires into a KSampler or the sampler of your choice. You'll also need a CLIP and a VAE - CRT ships ERNIE CLIP and ERNIE VAE AutoDL nodes for exactly that, so a complete ERNIE text-to-image stack is three nodes, all auto-downloading.

Installing it

It's part of CRT-Nodes. ComfyUI Manager → search CRT-Nodes, or:

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

Restart after. The requirements are heavy (opencv, transformers, librosa, faster-whisper…), so expect a slow install regardless of whether you use one node or all of them.

Common issues

• First run is slow - it's downloading the model (~8 GB for fp8) with a progress bar in the console. Subsequent runs load from disk.
• Model already exists - ensure_model checks the target path before downloading, so re-running won't refetch. If you want to force a redownload, delete the file from diffusion_models.
• ERNIE needs its matching CLIP - the Turbo text encoder is a Ministral 3B, not a Qwen. Using a generic CLIP will give you garbage or errors; use CRT's ERNIE CLIP.
• Red node / NaN after an update - right-click → Fix node (recreate); the AutoDL nodes have changed inputs across versions.