Manual Model Setup and Multiple LORA

[Engrish]

I checked out the project and tried to run it the first time. I noticed it is trying to download the default WAN 2.2 I2V models from a remote repo. However, the speed is extremely slow so I would like to download the models separately and place into the project folder. Do you have any instructions on how to do that? For example, where exactly they need to be placed (folder structure) and changes needed to the inference.py file? Also, if I want to add different LORA to the video generation, how do I do that with command line? From my understanding, it only accepts the SVI high and low LORA and nothing else.

Thank you

[wymanCV]

Hi, thank you for your interest. I just updated the README to list the required models. You can manually download them and organize them as suggested.

For LoRA merging, you can give it a try following this issue by simply modifying the load_lora in inference.py, as this codebase inherits all the functionalities provided by the official DiffSynth 2.0. But we havenot tried that.

Thank you.

[Engrish]

Hi wymanCV, thank you for the quick response. For now I'm going to skip adding more LORA so I can focus getting it up and running. I tried defining a new Model Config block inside inference.py.

model_configs=[ ModelConfig(model_id="Wan_Remix", origin_file_pattern="high/*.safetensors", offload_device="cpu"), ModelConfig(model_id="Wan_Remix", origin_file_pattern="low/*.safetensors", offload_device="cpu"), ModelConfig(model_id="DiffSynth-Studio/Wan-Series-Converted-Safetensors", origin_file_pattern="wan_umt5-xxl_bf16.safetensors", offload_device="cpu"), ModelConfig(model_id="DiffSynth-Studio/Wan-Series-Converted-Safetensors", origin_file_pattern="wan_2.1_vae.safetensors", offload_device="cpu"), ],

So I substituted the default models with Wan2.2 Remix I got from Civitai and placed them inside the models/Wan_Remix. However, when I run the command, it is still trying to download the models.

Traceback (most recent call last): File "c:\tools\Stable-Video-Infinity\inference.py", line 249, in <module> main() File "c:\tools\Stable-Video-Infinity\inference.py", line 235, in main processor = StreamingVideoProcessor(lora_path_high=args.lora_path_high, lora_path_low=args.lora_path_low, use_anchor=True, seed_multiplier=args.seed_multiplier) File "c:\tools\Stable-Video-Infinity\inference.py", line 16, in __init__ self.initialize_pipeline() File "c:\tools\Stable-Video-Infinity\inference.py", line 30, in initialize_pipeline self.pipe = WanVideoSviPipeline.from_pretrained( File "c:\tools\Stable-Video-Infinity\diffsynth\pipelines\wan_video_svi.py", line 130, in from_pretrained model_pool = pipe.download_and_load_models(model_configs, vram_limit) File "c:\tools\Stable-Video-Infinity\diffsynth\diffusion\base_pipeline.py", line 283, in download_and_load_models model_config.download_if_necessary() File "c:\tools\Stable-Video-Infinity\diffsynth\core\loader\config.py", line 99, in download_if_necessary self.download() File "c:\tools\Stable-Video-Infinity\diffsynth\core\loader\config.py", line 65, in download snapshot_download( File "C:\Users\Daniel\miniforge3\envs\svi_wan22\lib\site-packages\modelscope\hub\snapshot_download.py", line 132, in snapshot_download return _snapshot_download( File "C:\Users\Daniel\miniforge3\envs\svi_wan22\lib\site-packages\modelscope\hub\snapshot_download.py", line 300, in _snapshot_download endpoint = _api.get_endpoint_for_read( File "C:\Users\Daniel\miniforge3\envs\svi_wan22\lib\site-packages\modelscope\hub\api.py", line 522, in get_endpoint_for_read if not self.repo_exists( File "C:\Users\Daniel\miniforge3\envs\svi_wan22\lib\site-packages\modelscope\hub\api.py", line 670, in repo_exists raise Exception('Invalid repo_id: %s, must be of format namespace/name' % repo_type) Exception: Invalid repo_id: model, must be of format namespace/name

[wymanCV]

Hi, it seems that the issue is also related to model downloading.

I just found that DiffSynth supports downloading models from HuggingFace, see here. Maybe you could keep our original inference pipeline unchanged and simply switch the download source from ModelScope to HuggingFace.

[Engrish]

Related question, why do I have to re-download the files I already have on my computer? I was using ComfyUI for playing around with WAN so I already have the model files. I just want to copy them over to the project's models folder. Given that other more talented people are working on the ComfyUI workflows, I'll wait for that to come out.

[wymanCV]

Please check our pinned issue for details!

We’ve just released the latest SVI 2.0 Pro, which now supports a native ComfyUI workflow. It can generate both single long-shot videos and multi-shot videos with scene transitions, while maintaining consistent character ID. You can flexibly control these modes via text prompts.

You can also find lots of cool user-created workflows in the comments under that issue. We will eventually integrate and update them into our repo. Hope you enjoy this version!

Thank you!