Bedovyy

/

ERNIE-Image-Quantized

like 8

Have you tried original BF16 and does it works properly?
Could you let me know information about your GPU, ComfyUI version (and CUDA version if your GPU is nVidia's)?

I have tested latest version of ComfyUI by git pull, by the way.

the NVFP4 version works flawlessly with the official comfy wf. (But you have to disable the prompt enhancer; it's badly implemented. Somebody knows what model the enhancer is based on?). Thanks for your quants ;-)

PE is based on Mistral3-3B, the same as text encoder.
https://huggingface.co/baidu/ERNIE-Image/blob/main/pe/config.json

Nice work！I try to use the follow configuration，and just replace the ernie-image-turbo.safetensors with your fp8 safetensors，but it only generates noisy images。Could you please help me？
https://huggingface.co/Comfy-Org/ERNIE-Image
📂 ComfyUI/
├── 📂 models/
│ ├── 📂 diffusion_models/
│ │ ├── ernie-image.safetensors
│ │ └── ernie-image-turbo.safetensors
│ ├── 📂 text_encoders/
│ │ ├── ernie-image-prompt-enhancer.safetensors
│ │ └── ministral-3-3b.safetensors
│ └── 📂 vae/
│ └── flux2-vae.safetensors

you may misuse PE fp8 as clip? I got noisy images when doing this. ernie-image-turbo-fp8.safetensors works well on my device