[bug] Converting from diffusers format to ckpt format, but found that the model had no effect

[xdobetter]

I used zero123-hf project fine-tune zero123 105000.ckpt in my custom data, and then used this file conver_diffusers_to_zero123.py to convert the trained model diffusers format to ckpt, but I found that the converted ckpt model had no effect. Is this what I did wrong?

convert command

python conver_diffusers_to_zero123.py --checkpoint_path ../zero123/105000-train.ckpt --template_checkpoint_path ../zero123/temp/105000-temp.ckpt --original_config_file ../configs/sd-objaverse-finetune-c_concat-256.yaml --model_path ../outputs/scene_18

The 105000-temp.ckpt file is the official 105000.ckpt file that I renamed. The ../outputs/scene_18 is the location of my trained diffusers model.

The log trained using zero123-hf is as follows. You can see that the model is effective.

[xdobetter]

What I mean by having no effect is that when I used the model of zero123-hf fine-tune 105000.ckpt, I found that it had no effect on the official gradio_new.py on zero123, but it had effect on zero123-hf gradio_new.py .

In addition, when I lifted it to 3D on threestudio zero123, I also found that it had no effect.

I have verified diffusers mat -> ckpt mat -> diffusers mat to determine if there is a problem with your convert script. I found that it can use zero123-hf grario_new.py for infer, and it works.

Have you encountered these situations in your works? Thank you very much for giving me some suggestions.