GGUF and INT8 ConvRot and INT4 W4A8 ConvRot quantizations of the UNET for ERNIE-Image: https://huggingface.co/baidu/ERNIE-Image
NOTE: I did not make this model and I am not associated with it, I just quantized it using my CLI tool, ggufy.
The text encoder is also available in INT8 or W4A8, listed below the Turbo model.
The original BF16 text encoder and the VAE are also available from ComfyUI's hugging face: https://huggingface.co/Comfy-Org/ERNIE-Image
(Note: Personally I had pretty poor results from the "prompt enhancer" model, and don't recommend using it.)
PLEASE NOTE!
If you are using the GGUF versions, ComfyUI-GGUF does NOT support this model yet! However, you can simply add "ernie" to the list of supported models and it will work. This is on line 12 of the file custom_nodes/ComfyUI-GGUF/loader.py and when you are done, it should look like this:
MG_ARCH_LIST = {"flux", "sd1", "sdxl", "sd3", "aura", "hidream", "cosmos", "ltxv", "hyvid", "wan", "lumina2", "qwen_image", "ernie"}
Recommended Settings:
Sampler: euler
Schedule: simple
Steps: 20-50 (base), 8 (turbo)
CFG: 2-4 (base), 1 (turbo)
Description
FAQ
Comments (12)
i actually got this working with comfy-gguf, but i had to alter the loader.oy code with google gemini. took a few days to get it right, but i finally got it working and with a gguf of the text encoder, too. the prompt enchancer i left alone and it takes like 3 times longer (i'm runnin with a pathetic 4GB VRAM), but i get no OOM errors either way. Quality is not bad either and it has yet to make a spelling error or spit out any fake AI GibberGabber language. I can make a copy available to download on github if you are interested.
That's interesting, on my setup I only had to add it as an option and it worked fine. Are you using the original from this repo https://github.com/city96/ComfyUI-GGUF or a fork? Is your ComfyUI and GGUF plugin up to date?
Thanks compatible sdxl
Not working. Ernie text encoder incompatible with Comfy-GGUF.
And this model incompatible with GGUF-main extension which required for ernie text encoder.
Sorry you're having a hard time, it works fine for me.
@ferretduck do you use gguf text encoder?
@velanteg I have not tried it with a gguf text encoder, no
Thanks for sharing the GGUF version.
I’d be glad to see an INT4/INT8 (W4A4) version if possible.
I like Ernie Image’s prompt accuracy — the result usually matches expectations. At certain points, the accuracy is even excessive, since it requires a detailed description, including a negative prompt to fix anatomy, but I like how the text encoder works. In some cases, the text encoder even surprised me with how it handles understanding very complex prompts that, based on my generation results, I’m not sure other models would have been able to fulfill.
It’s a pity the community hasn’t yet unlocked this model’s potential
Your wish is my command.
Also added the text encoder as INT8 and W4A8, it's listed under the Turbo model.
Note that I did INT8 ConvRot and INT4 Convrot W4A8, not W4A4, as the W4A8 one is barely larger than W4A4 but is much better quality.
@ferretduck Thank you for responding so quickly.
I want to try the W4A8 format, the only problem is that I couldn’t previously find a solution for running W4A8 in ComfyUI - even the internet didn’t help me. The only thing I know is the W4A4 (int4-fast) loader, which refuses to run the A8 format.
Do you know of any way to run it in ComfyUI, or do you use a different interface for your work?
This is the only reason I usually ran W4A4, but if I found a way to run W4A8 inside Comfy, I would gladly try it
@solmp1 ComfyUI has native support for INT8, INT4, W4A8 and has for some time, it's part of ComfyUI Kitchen which is installed by default. You should be able to run it fine!