How to run ERNIE-Image locally.

Base or Turbo.

ERNIE-Image is Baidu’s 8B single-stream diffusion transformer, public since April 2026. It shares the Flux.2 VAE and latent, and it’s good at posters and text. Both versions use the same files around the model.

• ERNIE-Image is the full model. Baidu runs it at 50 steps and CFG 4, and it follows a negative prompt.
• ERNIE-Image-Turbo is distilled for 8 steps at CFG 1, with no negative prompt. It’s the faster place to start.

Files you need.

Comfy-Org’s repack has everything in bf16. There’s no fp8 or int8 version from Comfy-Org.

• Model ernie-image-turbo.safetensors ComfyUI/models/diffusion_models/ or ernie-image.safetensors, the full model, 16.1 GB
  16.1 GB Download
• Text encoder ministral-3-3b.safetensors ComfyUI/models/text_encoders/
  7.7 GB Download
• VAE flux2-vae.safetensors ComfyUI/models/vae/
  0.3 GB Download
• Optional ernie-image-prompt-enhancer.safetensors ComfyUI/models/text_encoders/ a language model that rewrites prompts. Not a second encoder
  6.9 GB Download

models/
├── diffusion_models/
│   └── ernie-image-turbo.safetensors
├── text_encoders/
│   ├── ministral-3-3b.safetensors
│   └── ernie-image-prompt-enhancer.safetensors   (optional)
└── vae/
    └── flux2-vae.safetensors

Smaller files: GGUF

unsloth publishes both versions as GGUF, built with ComfyUI-GGUF’s tools. Load them with the ComfyUI-GGUF Unet Loader. The Ministral 3 encoder has no GGUF that ComfyUI-GGUF reads yet, so it stays at 7.7 GB.

Q8_0 is close to the original. Q5_K_M and Q6_K are the usual middle ground.

What fits your computer.

At 1024 × 1024. The encoder runs first and ComfyUI moves it out before sampling, so the model file sets the limit. Only Baidu’s 24 GB figure is official; the other rows follow from the file sizes.

• 8 GBTight

  GGUF Q4_K_M (5.0 GB). The 7.7 GB encoder partly waits in system RAM.

• 12 GBFits

  GGUF Q5_K_M or Q6_K (5.9 to 6.8 GB).

• 16 GBFits

  GGUF Q8_0 (8.7 GB). The bf16 file is 16.1 GB and offloads part of itself.

• 24 GBFits

  The bf16 files, as Baidu intends.

• RTX 20 and olderSlow

  ComfyUI runs ERNIE-Image in bf16 or fp32 only. Cards without bf16 fall back to fp32, which is slower and needs more memory.

• MacUntested

  No ComfyUI reports on Apple Silicon yet. The bf16 files avoid the fp8 problems Macs have with other models.

• 8 GB graphics cards
• 12 GB cards
• 16 GB cards
• 24 GB cards

Set it up.

1. 1

   Update ComfyUI

   ERNIE-Image needs a ComfyUI from mid-April 2026 or later, with EmptyFlux2LatentImage and the ERNIE text encoder. In a manual install:

   Terminal, in the ComfyUI folder Copy

   git pull
   pip install -r requirements.txt

2. 2

   Download the files

   Model, encoder and VAE from the list above. Leave the prompt enhancer out if you don’t want it.

3. 3

   Open the template

   Pick Ernie Image Turbo: Text To Image or Ernie Image: Text to Image. Load CLIP gets ministral-3-3b with type flux2.

4. 4

   Add the shift

   The template has no shift node, so ComfyUI uses its default of 3. Baidu trained and runs with 4. Put a ModelSamplingSD3 node between Load Diffusion Model and the KSampler and set shift to 4.

5. 5

   Set the steps

   Turbo stays at 8 steps and CFG 1. For the full model, the template uses 20 steps and notes “Try 50 steps” for the original setting; Baidu’s card uses 50.

6. 6

   Skip or keep the enhancer, then run

   If the template’s enhancer nodes are active and you’d rather not use them, select them and press Ctrl+B to bypass them, so your prompt goes straight to the encoder.

Settings that work.

ERNIE-Image

Steps

50template: 20

CFG

4

Sampler

euler

Scheduler

simple

Shift

4ModelSamplingSD3. Default: 3

Size

1024 × 1024

Negative

yes

Latent

EmptyFlux2LatentImage

ERNIE-Image-Turbo

Steps

8

CFG

1

Shift

4

Negative

none

Turbo uses the same sampler, scheduler and latent. Baidu lists these sizes: 1024 × 1024, 848 × 1264, 1264 × 848, 768 × 1376, 1376 × 768, 896 × 1200 and 1200 × 896. When people complain about anatomy, the usual advice is one of Baidu’s sizes, the full step count and the enhancer.

The prompt enhancer.

The enhancer is a separate 3B language model. The template runs it through TextGenerate to expand a short prompt into a detailed one, then encodes the result with Ministral 3 as usual. Baidu’s own examples use it.

When it goes wrong.

“Dual clips not working” after loading the enhancer as a second encoder

The enhancer isn’t an encoder. Use one Load CLIP with ministral-3-3b, type flux2, and run the enhancer through TextGenerate or not at all.

A stall of several minutes at “model initializing” after the enhancer

The enhancer is still in memory. Add an unload or garbage-collect node between the enhancer and the sampler.

Missing EmptyFlux2LatentImage or an unknown text encoder type

ComfyUI is too old for ERNIE-Image. Update it.

Results look off next to Baidu’s examples

Check the shift is 4, not the default 3, and that the full model runs 50 steps at one of Baidu’s sizes.

Out of memory with the bf16 file on 16 GB

Use a GGUF build: Q8_0 is 8.7 GB and close to the original.

Questions.

Sources: ERNIE-Image model card, ERNIE-Image-Turbo model card, scheduler config (shift 4), ComfyUI ERNIE-Image tutorial, Comfy-Org ERNIE-Image files, dual clip thread, enhancer stall thread, unsloth GGUF.