bad results, how to solve?

[qmzpg]

EulerAncestralDiscrete_0_gs7.5_s100.mp4

The command is:
CUDA_VISIBLE_DEVICES=0,1,2,3 torchrun --nnodes=1 --nproc_per_node 1 --master_port 29514
-m opensora.sample.sample
--model_type "inpaint"
--model_path "/home//llm/models/Open-Sora-Plan-v1.3.0/any93x640x640_i2v"
--version v1_3
--num_frames 33
--height 352
--width 640
--max_hxw 236544
--crop_for_hw
--cache_dir "../cache_dir"
--text_encoder_name_1 "/home//llm/models/mt5-xxl"
--text_prompt examples/cond_prompt.txt
--conditional_pixel_values_path examples/cond_pix_path.txt
--ae WFVAEModel_D8_4x8x8
--ae_path "/home/*******/llm/models/Open-Sora-Plan-v1.3.0/vae"
--save_img_path "./save_path"
--fps 6
--guidance_scale 7.5
--num_sampling_steps 100
--max_sequence_length 512
--sample_method EulerAncestralDiscrete
--seed 1234
--num_samples_per_prompt 1
--rescale_betas_zero_snr
--prediction_type "v_prediction"
--noise_strength 0.0 \

[yunyangge]

Due to the data used during training, selecting num_frames=93 will yield relatively better results. Additionally, because of the current model's limited capability, inference in multi-person scenarios has a higher failure rate. This can be mitigated by adjusting the seed and prompt multiple times to achieve the best results. We are working on training a new version of the Open-Sora Plan model to provide better generation quality. Thank you for your understanding.

[WangRongsheng]

how to set --text_prompt examples/cond_prompt.txt?