search

kohya-ss / sd-scripts

Apache License 2.0
4.54k stars 770 forks source link

is the sd3_minimal_inference something wrong？ #1388

Open lzran opened 3 weeks ago

lzran commented 3 weeks ago

Thanks for this update for sd3 generate image，but I used sd3_minimal_inference.py to generate image prompt like “a cat holding a sign that says hello world”，it outputs only a cat，it seems like the text has been truncated，somebody knows what's going on？

kohya-ss commented 3 weeks ago

One of Text Encoders might be missing. Please use sd3_medium_incl_clips_t5xxlfp16.safetensors or specify each .safetensors file for --clip_g, --clip_l and --t5xxl options.

lzran commented 2 weeks ago

One of Text Encoders might be missing. Please use sd3_medium_incl_clips_t5xxlfp16.safetensors or specify each .safetensors file for --clip_g, --clip_l and --t5xxl options.

but if use the sd3_medium_incl_clips_t5xxlfp16.safetensors seems need to download a big model named google/t5-v1_1-xxl and its too huge(100G), wating your updating for sd3_medium.safetensors

kohya-ss commented 2 weeks ago

sd3_medium_incl_clips_t5xxlfp16.safetensors is 15.8GB. This file will work. https://huggingface.co/stabilityai/stable-diffusion-3-medium/blob/main/sd3_medium_incl_clips_t5xxlfp16.safetensors

lzran commented 2 weeks ago

sd3_medium_incl_clips_t5xxlfp16.safetensors is 15.8GB. This file will work. https://huggingface.co/stabilityai/stable-diffusion-3-medium/blob/main/sd3_medium_incl_clips_t5xxlfp16.safetensors

I used this model sd3_medium_incl_clips_t5xxlfp16.safetensors seems the result is the same, if I specify .safetensors file for --t5xxl options., the sd3_models.py prompts that google/t5-v1_1-xxl is needed.

kohya-ss commented 2 weeks ago

Do you specify t5xxl_fp16.safetensors for --t5xxl option? The t5xxl tokenizer is downloaded, but the model is not downloaded in my env.

lzran commented 2 weeks ago

Do you specify t5xxl_fp16.safetensors for --t5xxl option? The t5xxl tokenizer is downloaded, but the model is not downloaded in my env.

Whether or not I specify t5xxl_fp16.safetensors for --t5xxl option, sd3_models.py comes out the same error: can't load tokenizer for 'google/t5-v1_1-xxl'.the commands excuted is "python sd3_minimal_inference.py --ckpt_path /workspace/sd3/model/d3_medium_incl_clips_t5xxlfp16.safetensors --t5xxl /workspace/sd3/text_encoders/t5xxl_fp16.safetensors --output_dir /workspace/imgs/test"