ERNIE_Turbo NVFP4 Model (CRT AutoDL)

Same idea as the plain ERNIE Model loader - no file path, no dropdown, it just fetches the weights and hands you a MODEL - but this one grabs the distilled Turbo checkpoint in NVFP4 format specifically. That combination (Turbo + NVFP4) is worth understanding before you drop this node in your graph, because it comes with a real hardware requirement the node itself won't warn you about.

What you're actually downloading

ERNIE-Image-Turbo is Baidu's DMD-plus-RL distilled variant of ERNIE-Image - roughly six times faster than the full SFT checkpoint at comparable quality, running at 8 steps and guidance 1.0 versus the base model's 50 steps at guidance 4.0. It's the variant nearly every community post about ERNIE actually used. Same strengths and weaknesses as the base model apply: genuinely good at structured layout and text-in-image, weaker at photorealism than Z-Image or Flux 2 Klein, and carrying the same documented ethnic-bias and grid-artifact issues.

NVFP4 is the quantization format layered on top, and this is the part that actually gates whether this node is useful to you: NVFP4 is a Blackwell-only format - it needs an RTX 50-series card. On the model this format is best documented for (Flux 2 Klein), the payoff is real: BFL claims up to 2.7x faster inference and 55% less VRAM on 50-series hardware versus fp8. Baidu's own ERNIE NVFP4 quantization isn't separately documented in the same depth, but the hardware constraint is the same either way - this isn't a "smaller file, runs anywhere" quantization like GGUF, it's a specific-hardware speed format.

The inputs and outputs that matter

Same three optimization toggles as the plain ERNIE Model loader:

• patch_cublaslinear (default false) - routes matmuls through cuBLAS for speed, the loader-level equivalent of ComfyUI's --fast cublas_ops flag.
• sage_attention (default auto) - picks or forces a SageAttention kernel (disabled, auto, or one of several named kernel variants like sageattn3). Leave it on auto unless you specifically know which kernel your GPU needs.
• enable_fp16_accumulation (default true) - fp16 matmul accumulation for a speed gain on GPUs that support it.

Output: a single MODEL.

Installing it

Same install as the rest of the pack - this node isn't a separate download, it ships inside CRT-Nodes:

ComfyUI Manager - search CRT-Nodes, install, restart.

Manual:

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

Restart ComfyUI. The actual model weights fetch on first run of the node, not at pack-install time - this specific AutoDL family isn't itemized in the CRT-Nodes README's own node catalog table, so treat the download-on-first-use behavior as something to expect rather than something the docs walk you through in detail.

Common issues

• Not a 50-series card? This node isn't for you yet. NVFP4 is a Blackwell-specific format. On anything older (40-series, 30-series and below), use the plain ERNIE Model (CRT AutoDL) node or an fp8/GGUF build instead - trying to run an NVFP4 checkpoint on unsupported hardware is the single most likely source of a hard failure here, not a config mistake.
• Slow or stuck on first run - that's the download, not inference. NVFP4 files are still multi-gigabyte; give it time before assuming something's broken.
• Output looks worse than expected for "Turbo" - remember Turbo trades quality for the 6x-ish speedup via distillation; that's the checkpoint working as intended, not a quantization artifact. If quality matters more than speed, the plain ERNIE Model node (the SFT checkpoint) is the one to reach for instead.
• sage_attention errors on a specific kernel choice - set it back to auto, which is written to pick something your GPU actually supports rather than forcing a mismatch.