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, 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, 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 = best open-weight text rendering + layout T2I. Qwen-Image-Edit = instruction editing. Flux Kontext = reference editing. Z-Image Turbo = 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 — 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 GPUs <8 GB · Q6_K 8–12 GB · Q8_0 12–16 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, ernie-anime-v1.safetensors — 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) — verify which pack provides it via ComfyUI-Manager if it shows as missing (unverified pack origin).

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

Prompt enhancer path (optional): 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.

Image-to-image (refine) path — 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 (0.35–0.4) to refine/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.

Combo pipelines (ERNIE↔ZIT): 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.

Settings (extracted from the shipped KSamplers)

• ModelSamplingAuraFlow shift = 3.1 is applied to the ERNIE model before sampling (flow-matching shift). Keep it.
• CFG = 1 for Turbo → negative conditioning is effectively inert; the graph still wires a ConditioningZeroOut as the negative.
• Resolution: shipped latent is 1920×1088 (EmptySD3LatentImage). ERNIE is a high-res-capable DiT; 1024–2048 on the long edge is reasonable. Use EmptySD3LatentImage for ERNIE latents.
• Base (non-Turbo) ernie-image: bump steps to ~50 and raise cfg (e.g. 3.5–5) since it is not distilled.

Prompt / instruction style

ERNIE rewards descriptive, structured natural-language prompts, and is unusually strong at literal text rendering. Write the exact text you want to appear in quotes.

A vintage travel poster of Kyoto in autumn, bold title text reading "KYOTO" at the top,
maple leaves, Mount fuji silhouette, clean vector layout, muted warm palette

A 3-panel manga page: panel 1 a samurai drawing his sword, panel 2 close-up of his eyes,
panel 3 a wide shot of cherry blossoms falling, black-and-white ink, speech bubble "参る"

• For typography/signage: state the literal string ("a neon sign that says 'OPEN'"), placement, and font feel.
• For layout: name the panel/grid structure and what goes in each region.
• Multilingual prompts (incl. Chinese) work — the built-in enhancer's system prompt is Chinese.
• This is txt2img phrasing, not edit phrasing. Do not write "change the…/remove the…" expecting grounded edits.

Complete API-format workflow (ERNIE-Image-Turbo text-to-image)

Derived from the source graph, flattened to API format (no subgraphs/virtual wires). Enhancer omitted for clarity — CLIPTextEncode takes the prompt directly.

{
  "1": { "class_type": "UnetLoaderGGUF", "inputs": { "unet_name": "ernie-image-turbo-Q8_0.gguf" } },
  "2": { "class_type": "CLIPLoader", "inputs": { "clip_name": "ministral-3-3b.safetensors", "type": "flux2", "device": "default" } },
  "3": { "class_type": "VAELoader", "inputs": { "vae_name": "flux2-vae.safetensors" } },
  "4": { "class_type": "ModelSamplingAuraFlow", "inputs": { "model": ["1", 0], "shift": 3.1 } },
  "5": { "class_type": "CLIPTextEncode", "inputs": { "clip": ["2", 0], "text": "A vintage travel poster of Kyoto in autumn, bold title text reading \"KYOTO\" at the top, maple leaves, clean vector layout, muted warm palette" } },
  "6": { "class_type": "ConditioningZeroOut", "inputs": { "conditioning": ["5", 0] } },
  "7": { "class_type": "EmptySD3LatentImage", "inputs": { "width": 1920, "height": 1088, "batch_size": 1 } },
  "8": { "class_type": "KSampler", "inputs": {
    "model": ["4", 0], "positive": ["5", 0], "negative": ["6", 0], "latent_image": ["7", 0],
    "seed": 997032332094579, "steps": 9, "cfg": 1, "sampler_name": "euler", "scheduler": "simple", "denoise": 1
  } },
  "9": { "class_type": "VAEDecode", "inputs": { "samples": ["8", 0], "vae": ["3", 0] } },
  "10": { "class_type": "SaveImage", "inputs": { "images": ["9", 0], "filename_prefix": "ernie_image" } }
}

Image-to-image (refine) variant

Replace the empty latent with an encoded source image and lower denoise. This restyles/refines a single image; it is not instruction editing.

{
  "11": { "class_type": "LoadImage", "inputs": { "image": "source.png" } },
  "12": { "class_type": "ImageResize+", "inputs": { "image": ["11", 0], "width": 1024, "height": 1024, "interpolation": "lanczos", "method": "keep proportion", "condition": "always", "multiple_of": 0 } },
  "13": { "class_type": "VAEEncode", "inputs": { "pixels": ["12", 0], "vae": ["3", 0] } }
}

Then in the KSampler set "latent_image": ["13", 0] and "denoise": 0.4.

Adding LoRAs

Insert a Power Lora Loader (rgthree) between the UNet loader and ModelSamplingAuraFlow (model: ["1",0] → loader → ["4"].model). In API format you can substitute LoraLoaderModelOnly with lora_name: "ernie-anime-v1.safetensors", strength_model: 0.5.

VRAM

• ERNIE-Image-Turbo GGUF: Q5_K_S <8 GB · Q6_K 8–12 GB · Q8_0 12–16 GB+ (installer's own guidance).
• Ministral-3-3B encoder + Flux2 VAE add a few GB. The graph includes easy cleanGpuUsed / easy clearCacheAll nodes between stages — keep them for the combo/two-pass pipelines so VRAM is freed before swapping models.
• Running the ERNIE↔ZIT combos loads two UNets; budget for both or run the single-model ERNIE group only.

Troubleshooting

• UnetLoaderGGUF / CLIPLoaderGGUF missing → install ComfyUI-GGUF (city96).
• Power Lora Loader / Image Comparer / Label missing → install rgthree-comfy.
• ImageResize+ missing → install ComfyUI_essentials.
• TextGenerate missing (prompt enhancer) → install via ComfyUI-Manager search; pack origin unverified. If unavailable, just set the ComfySwitchNode to use the raw prompt (switch=false) and skip enhancement.
• CLIP type error on ministral → ensure CLIPLoader type is flux2 (not qwen_image/lumina2). The lumina2 type belongs to the Z-Image (Qwen3) encoder, not ERNIE.
• Wrong VAE artifacts → ERNIE must use flux2-vae.safetensors; ae.safetensors is the Z-Image VAE.
• Blurry / undercooked output → confirm ModelSamplingAuraFlow shift=3.1 is wired and steps ≥8 for Turbo; for base ernie-image use ~50 steps + higher cfg.
• You wanted to EDIT a photo and it ignored the instruction → expected. ERNIE is txt2img; use the qwen-image-edit skill or Flux Kontext for grounded edits.
• Installer pulled Z-Image files too → expected (the scripts are Z-Image-derived). Harmless; those files only feed the combo pipelines.

Tips

1. Lead with the literal text you want rendered, in quotes — that's ERNIE's headline strength.
2. Use the prompt enhancer for short/lazy prompts; turn it off (ComfySwitchNode false) when you've written a detailed prompt yourself.
3. Use analyze_workflow before executing the shipped graph — it has dozens of group-boxed variants gated by Fast Groups Bypasser (rgthree); the analyzer summary is far easier than reading raw JSON.
4. Most groups are bypassed (mode 4) by default in the source file — enable only the pipeline you want via the group bypasser, or build the clean API workflow above.
5. To choose a model: ERNIE = text/layout T2I, Z-Image Turbo = fast general T2I, Qwen-Image-Edit / Flux Kontext = actual editing.