JettHu
JettHu
It looks like using `--fp32-text-enc` affects the results. [refer to](https://github.com/TencentQQGYLab/ELLA?tab=readme-ov-file#4-flant5-must-run-in-fp16-mode) The results on my machine are similar to yours.
> It looks like using `--fp32-text-enc` affects the results. [refer to](https://github.com/TencentQQGYLab/ELLA?tab=readme-ov-file#4-flant5-must-run-in-fp16-mode) > > The results on my machine are similar to yours. > > The effect is somewhat different on...
It seems that you may have used the sdxl model. Currently, the open source ella only has sd1.5. [refer to](https://github.com/TencentQQGYLab/ELLA/issues/13#issuecomment-1998834009)
Try to concat the `CONDITIONING` of `CLIP Text Encode` with the `ELLA` `CONDITIONING` like in the [lora example](https://github.com/TencentQQGYLab/ComfyUI-ELLA/blob/main/examples/workflow_lora.png).
I mean, use CLIP to encode "nfsw" at the same time, and then concat the two results. Our training set lacks the word "nfsw" and pictures, so that need to...
Not tested yet. But already in the plan.
Sorry, I've been busy lately, so this has a lower priority.
have you try update your comfyui version
Check your comfyui code. There is this method in the [comfyui repo](https://github.com/comfyanonymous/ComfyUI/blob/master/comfy/model_patcher.py#L164)
I just tried it on comfyui and the problem did not occur. (comfyui commit-id: 27d5808fc491c7174abc6f407e7dc11c6a7a1ec0) It seems that no one else has encountered it either. It may be damaged by...