Ernie Image

ERNIE-Image / ERNIE-Image-Turbo Workflows

What this is (read first)

ERNIE-Image is Baidu's open-weight TEXT-TO-IMAGE model, an ~8B single-stream Diffusion Transformer (DiT), Apache-2.0, released April 2026 and repackaged for ComfyUI by Comfy-Org. It is not an instruction-based image editor.

• ERNIE-Image (base): ~50 steps for peak quality.
• ERNIE-Image-Turbo: distilled (Distribution Matching Distillation + RL), high-fidelity in ~8 steps at cfg 1. The downloaded pack uses Turbo (ernie-image-turbo-*.gguf).

Pick ERNIE when the job is precise text/typography rendering (multilingual, including Chinese), posters/signage/UI mockups, manga/anime storyboards and multi-panel layouts, or structured multi-object scenes from a complex prompt. Do not pick ERNIE for "edit this photo / change the shirt / swap the background". That is instruction-grounded editing, which ERNIE does not do. Use qwen-image-edit or Flux Kontext for those. ERNIE's "image-to-image" here is plain denoise-based refinement (style pass / detail pass), not reference-grounded editing.

Niche vs siblings. ERNIE is the best open-weight text rendering + layout T2I. Qwen-Image-Edit does instruction editing. Flux Kontext does reference editing. Z-Image Turbo does fast general T2I, and this same pack pairs the two; see Combo pipelines.

Separated packs (render-verified)

The original ernie monolith was a single toggle-template graph (every pipeline shipped bypassed; you activated one via the rgthree group toggles). It's now split into standalone, single-purpose packs, each a clean activated graph that renders headlessly with no group-toggling:

Working details verified live: the prompt-enhancer LLM is OFF by default (the ENHANCE PROMPT boolean is false; leave it off unless you want the 3B enhancer to rewrite the prompt). The grain/sharpen post-proc (FastFilmGrain/FastLaplacianSharpen, comfyui-vrgamedevgirl) needs librosa installed. In ernie-combo the Z-Image half's VAE is saved as z-image-ae.safetensors; its weights differ from Flux/ERNIE's ae.safetensors despite the same size, and the rename avoids a filename clash.

Source of truth & a provenance warning

This skill is derived from the actual pack files in C:\Users\Artokun\Downloads\:

• ERNIE-IMAGE-ULTRA-WORKFLOW.json (authoritative; the ComfyUI graph)
• ERNIE-IMAGE_ULTRA-MODELS-NODES_INSTALL.bat, ...-COMFYUI-MANAGER_AUTO_INSTALL.bat, ...-AUTO_INSTALL-RUNPOD.sh

Installer warning (verified). The three install scripts are copy-pasted from a Z-Image pack. Their headers literally say "Z-IMAGE-BASE"/"Z-IMAGE Base", and they download both ERNIE and Z-Image files. The model URLs/folders below are taken from those scripts but mirror this confusion. They pull z_image_turbo-*.gguf, Qwen3-4B-*.gguf, and ae.safetensors, which belong to the Z-Image half of the combo, not ERNIE. The ERNIE-only files are flagged below. All weights come from a third-party mirror huggingface.co/Aitrepreneur/FLX, not the official huggingface.co/Comfy-Org/ERNIE-Image (which hosts the same filenames; see Official sources).

Models

ERNIE-Image (the files ERNIE actually uses)

Confirmed from the workflow's virtual wires (Set_*/GetNode): the nodes tagged "ERNIE" resolve to these exact files.

Quant guidance from the installer: Q5_K_S for GPUs <8 GB · Q6_K for 8 to 12 GB · Q8_0 for 12 to 16 GB+.

Z-Image Turbo (bundled in the same pack — the "ZIT" half)

The workflow also wires a parallel Z-Image Turbo pipeline for ERNIE→ZIT / ZIT→ERNIE combos. These files are Z-Image's, not ERNIE's. Do not confuse them:

LoRAs (referenced in the Power Lora Loader, off by default)

hirohiko-araki-style-ERNIE_000001250.safetensors and ernie-anime-v1.safetensors are community ERNIE style LoRAs, loaded via Power Lora Loader (rgthree) (both toggled off in the shipped graph). Not in the installer; user-supplied.

Upscalers / post (shared)

4x-ClearRealityV1.pth, RealESRGAN_x4plus_anime_6B.pth → models/upscale_models/.

Installation

Custom nodes (git clone into ComfyUI/custom_nodes/)

All three installers clone the same set:

The graph also uses TextGenerate, TextBox1, StringReplace, ComfySwitchNode, PreviewAny, SetNode/GetNode, PrimitiveBoolean, ModelSamplingAuraFlow, ConditioningZeroOut, EmptySD3LatentImage, EmptyFlux2LatentImage. Most are builtin or come from the packs above. SetNode/GetNode are from KJNodes. TextGenerate (runs the prompt-enhancer LLM) has an unverified pack origin; verify which pack provides it via ComfyUI-Manager if it shows as missing.

Model downloads (exact URLs from the installer)

Base URL HF = https://huggingface.co/Aitrepreneur/FLX/resolve/main (third-party mirror). !MODEL_VERSION! ∈ {Q5_K_S, Q6_K, Q8_0}.

# ERNIE (the files ERNIE actually uses)
unet/ernie-image-turbo-<Q>.gguf                <HF>/ernie-image-turbo-<Q>.gguf?download=true
text_encoders/ministral-3-3b.safetensors       <HF>/ministral-3-3b.safetensors?download=true
text_encoders/ernie-image-prompt-enhancer.safetensors  <HF>/ernie-image-prompt-enhancer.safetensors?download=true
vae/flux2-vae.safetensors                       <HF>/flux2-vae.safetensors?download=true

# Z-Image half (bundled; only needed for the ZIT combo pipelines)
unet/z_image_turbo-<Q>.gguf                      <HF>/z_image_turbo-<Q>.gguf?download=true
text_encoders/Qwen3-4B-UD-Q6_K_XL.gguf          <HF>/Qwen3-4B-UD-Q6_K_XL.gguf?download=true
vae/ae.safetensors                               <HF>/ae.safetensors?download=true

# Upscalers
upscale_models/4x-ClearRealityV1.pth            <HF>/4x-ClearRealityV1.pth?download=true
upscale_models/RealESRGAN_x4plus_anime_6B.pth   <HF>/RealESRGAN_x4plus_anime_6B.pth?download=true

Official sources (prefer these over the mirror)

The same filenames are hosted officially at huggingface.co/Comfy-Org/ERNIE-Image (unet|diffusion_models/, text_encoders/, vae/). Apache-2.0. Original model: github.com/baidu/ERNIE-Image. Comfy day-0 docs: docs.comfy.org/tutorials/image/ernie-image/ernie-image. The official repo also ships non-GGUF ernie-image.safetensors / ernie-image-turbo.safetensors (load with UNETLoader instead of UnetLoaderGGUF).

How the pipeline works (at a glance)

The big graph is a menu of group-boxed pipelines built from the same blocks. Core ERNIE-Image text-to-image flow:

UnetLoaderGGUF (ernie-image-turbo) ──► Power Lora Loader (rgthree) ──► ModelSamplingAuraFlow (shift=3.1) ──► MODEL
CLIPLoader (ministral-3-3b, type=flux2) ──► CLIP ──► CLIPTextEncode (positive)
                                              └──► ConditioningZeroOut  ──► negative   (cfg=1, so negative ≈ unused)
VAELoader (flux2-vae) ──► VAE
EmptySD3LatentImage (1920×1088) ──► LATENT
        │
KSampler (steps≈8–9, cfg=1, euler, simple, denoise=1) ──► VAEDecode ──► SaveImage / post

In the optional prompt enhancer path, the short user prompt + {width}/{height} are templated into a Chinese system prompt, fed to TextGenerate (which runs ernie-image-prompt-enhancer), and a ComfySwitchNode chooses raw prompt (switch=false) vs. enhanced prompt (switch=true) before CLIPTextEncode.

The image-to-image (refine) path is not editing. The graph's "ERNIE IMAGE TO IMAGE" groups take a LoadImage → ImageResize+ (1024, keep proportion, lanczos) and VAEEncode it, then run KSampler at low denoise (0.35 to 0.4) to refine or restyle. This is a denoise pass over a single source image; it does not follow edit instructions.

About the ~5 LoadImage + ~5 VAEEncode nodes: they are not multi-reference compositing. Each LoadImage feeds a separate pipeline variant (single-image img2img, or the combo refine stages). One source image per pipeline. The extra VAEEncodes are the encode steps for those independent img2img / two-pass refine chains.

The combo pipelines (ERNIE↔ZIT), with group titles ERNIE ---> ZIT COMBO, ZIT ---> ERNIE COMBO, TWO TIMES COMBO ..., chain ERNIE and Z-Image Turbo as a two-pass generate→refine, with optional film-grain (FastFilmGrain) or sharpening (FastLaplacianSharpen) finishing and SIMPLE UPSCALE.