ERNIE Model (CRT AutoDL)

No file-path widget, no dropdown of checkpoints in your models/ folder - this node just fetches Baidu's ERNIE-Image diffusion backbone for you and hands you back a ready MODEL. It's one of CRT-Nodes' auto-download loaders: point-and-click access to a specific, named model instead of the usual "go find the weights, put them in the right folder, restart" ritual.

What ERNIE-Image actually is (and isn't)

ERNIE-Image is Baidu's second attempt at open text-to-image - an 8B DiT released under Apache 2.0 in April 2026. It landed with day-one ComfyUI support and a real, if narrow, strength: structured layout and text-in-image, the kind of thing infographics, multi-panel comics, and labelled diagrams need and most photoreal-focused models are bad at. It is not a photorealism model. Head-to-head comparisons the community ran consistently landed on Z-Image or Flux 2 Klein for a straight portrait, and ERNIE carries a documented ethnic bias strong enough that a community LoRA exists specifically to patch it, plus visible grid/noise artifacts in a lot of generations. If you're reaching for this node, know what you're reaching for it for - infographics and text-heavy layouts, or as a LoRA-training target where the reports (undertested, but consistent) say it converges unusually fast.

Reality check on adoption: mentions of the model roughly halved every month after release, and it never grew a LoRA ecosystem or a finetune library the way Z-Image or Klein did. This is a real, useful, permissively licensed model for a specific job - not the daily driver.

The inputs and outputs that matter

Three widgets, all performance/optimization toggles rather than anything about which checkpoint to fetch (that part is hardcoded to ERNIE-Image by the node itself):

• patch_cublaslinear (default false) - patches Linear layers to route through cuBLAS for faster matmuls, the same idea behind ComfyUI's built-in --fast cublas_ops launch flag, just exposed as a per-loader toggle here.
• sage_attention (default auto) - which attention kernel to use. Options run from disabled through auto to several specific SageAttention kernel variants (sageattn_qk_int8_pv_fp16_cuda, sageattn_qk_int8_pv_fp8_cuda, sageattn3, and others). SageAttention is a quantized-attention library that trades a sliver of precision for real speed; auto picks a kernel based on your GPU, and the named options let you force a specific one if you know your hardware supports it.
• enable_fp16_accumulation (default true) - another speed/precision trade, mirroring ComfyUI's fp16_accumulation fast-mode flag: lets matmul accumulation happen in fp16 rather than fp32 on GPUs that support it, for a speed gain at a small precision cost.

Output: a single MODEL.

Installing it

CRT-Nodes' README doesn't list this AutoDL family in its own node catalog table - it documents the pattern generically (built-in HTTP download utilities, API routes for model status/download) for the LTX 2.3 AutoDownload node, but the ERNIE/Flux2Klein/Krea2 loaders aren't individually itemized. Install the pack the normal way and this node comes with it:

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, then drop this node into your graph - the actual ERNIE-Image weights download automatically the first time you run it, not at install time. Budget for a multi-gigabyte download and a pause on first use.

The pack's author (CRYPT_EXE / PGCRT) has done this auto-download pattern before, in a separate node pack (CRT-HeartMuLa) that fetches its music-generation model on first use the same way - this isn't a one-off feature bolted on for ERNIE specifically, it's the pack's established approach to model distribution.

Common issues

• First run hangs or is very slow - that's the model download, not the sampler. There's no progress bar guarantee here; check your ComfyUI console for HTTP download activity before assuming it's stuck.
• sage_attention set to a specific kernel throws an error - the named kernels are tied to particular GPU generations. If you force one your card doesn't support, set it back to auto (which picks a compatible kernel) or disabled (plain attention, no speedup, but always works).
• Download fails silently or the node errors on a fresh install - this loader is fetching from wherever Baidu's weights are hosted; if that changes or is temporarily unavailable, there's no local fallback since there's no file-path widget to point at an already-downloaded copy. Check the CRT-Nodes GitHub issues or Discord if a download consistently fails.
• Wrong tool for the job - if you're chasing photorealism or a clean single-portrait render, this isn't the model; reach for Z-Image or Flux 2 Klein instead and save this loader for layout/infographic work or LoRA training experiments.