ERNIE CLIP (CRT AutoDL)

ERNIE-Image is Baidu's April 2026 entry into open text-to-image models - an 8B Apache 2.0 checkpoint that had, on paper, everything the community says it wants: no licensing catch, day-one ComfyUI support, no built-in NSFW filter. It still didn't build much of an ecosystem. Corpus mentions fell from 183 threads in its launch month to 78 by June, no finetunes appeared, no LoRA library formed, and its promised edit model never shipped. What did survive, and it's a real, specific pocket rather than a consolation prize: ERNIE is genuinely good at structured layout and text-in-image - infographics, multi-panel comics, posters, labelled diagrams - a niche nothing else open handles as well. This node sets up the text encoder side of that model.

What it does

Zero inputs, one output: CLIP. The naming and empty input signature put it in the pack's "AutoDL" family, the same pattern used for the pack's confirmed LTX 2.3 auto-downloader (built-in HTTP download plus API routes for status). The strong implication is that this node fetches ERNIE-Image's text encoder weights itself on first run rather than requiring you to hunt down and place the file manually - though the README doesn't spell out this node's specifics beyond its category label, so treat the auto-download behavior as a reasonable inference from the naming pattern rather than a documented guarantee.

ERNIE's text encoder is worth knowing about on its own terms: it's a Mistral3Model-based encoder, smaller than the Qwen or T5 encoders most competing models use, which is part of what lets Baidu claim a 24GB consumer-GPU target for the whole pipeline. It also carries an unexposed vision component, since the underlying architecture has multimodal roots even though the public interface is text-only.

The output that matters

Just CLIP, wired into whatever encode node your ERNIE-Image workflow uses, alongside an ERNIE UNETLoader and VAE. If you're pairing this with the built-in Prompt Enhancer some ERNIE workflows use - a separate small model that rewrites terse prompts before the backbone sees them - know that it's not a free win: Baidu's own benchmarks show it improving counting and positioning while trading away some attribute-binding precision, and several users report better likeness and coherence with it bypassed on prompts that are already specific.

How to install it

Part of CRT-Nodes. Via ComfyUI Manager, search CRT-Nodes and install. Manually: git clone https://github.com/PGCRT/CRT-Nodes.git into custom_nodes, pip install -r requirements.txt, restart. You're pulling the whole pack's dependency stack for this one auto-downloader, same as every other CRT-Nodes install.

Common issues & troubleshooting

Ethnic bias in generated faces. This is a documented, real limitation of the ERNIE-Image model itself, not this loader - community testing found the model strongly biased toward Asian faces even with nationality terms stacked in the prompt and the enhancer disabled. A community-published LoRA exists specifically to correct this if it matters for your use case.

Visible grid or noise patterns in output. Also a model-level issue reported independently across multiple release threads, not something this text encoder node can affect.

Output degrades noticeably above 1024x1024. ERNIE-Image was trained only up to that resolution - going higher is a known way to get mangled anatomy, more so than with comparable models. Keep generations at or below 1024x1024 and upscale separately if you need more resolution.

First run stalls. If the auto-download inference above is correct, that's the weight fetch happening on first execution, not a hang - subsequent runs should be fast once the file is cached.