> 📝 Click on the language section to expand / 蚀語をクリックしお展開 # Wan 2.1/2.2 ## Overview / 抂芁 This is an unofficial training and inference script for [Wan2.1](https://github.com/Wan-Video/Wan2.1) and [Wan2.2](https://github.com/Wan-Video/Wan2.2). The features are as follows. - fp8 support and memory reduction by block swap: Inference of a 720x1280x81frames videos with 24GB VRAM, training with 720x1280 images with 24GB VRAM - Inference without installing Flash attention (using PyTorch's scaled dot product attention) - Supports xformers (training and inference) and Sage attention (inference only) - Support for Wan2.2 model architecture, only for 14B models This feature is experimental.
日本語 [Wan2.1](https://github.com/Wan-Video/Wan2.1) および [Wan2.2](https://github.com/Wan-Video/Wan2.2) の非公匏の孊習および掚論スクリプトです。 以䞋の特城がありたす。 - fp8察応およびblock swapによる省メモリ化720x1280x81framesの動画を24GB VRAMで掚論可胜、720x1280の画像での孊習が24GB VRAMで可胜 - Flash attentionのむンストヌルなしでの実行PyTorchのscaled dot product attentionを䜿甚 - xformers孊習ず掚論およびSage attention掚論のみ察応 - Wan2.2モデルアヌキテクチャのサポヌト14Bモデルのみ この機胜は実隓的なものです。
## Download the model / モデルのダりンロヌド ### Wan2.1 Download the T5 `models_t5_umt5-xxl-enc-bf16.pth` and CLIP `models_clip_open-clip-xlm-roberta-large-vit-huge-14.pth` from the following page: https://huggingface.co/Wan-AI/Wan2.1-I2V-14B-720P/tree/main Download the VAE from the above page `Wan2.1_VAE.pth` or download `split_files/vae/wan_2.1_vae.safetensors` from the following page: https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/tree/main/split_files/vae Download the DiT weights from the following page: https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/tree/main/split_files/diffusion_models Wan2.1 Fun Control model weights can be downloaded from [here](https://huggingface.co/alibaba-pai/Wan2.1-Fun-14B-Control). Navigate to each weight page and download. The Fun Control model seems to support not only T2V but also I2V tasks. Please select the appropriate weights according to T2V, I2V, resolution, model size, etc. `fp16` and `bf16` models can be used, and `fp8_e4m3fn` models can be used if `--fp8` (or `--fp8_base`) is specified without specifying `--fp8_scaled`. **Please note that `fp8_scaled` models are not supported even with `--fp8_scaled`.** (Thanks to Comfy-Org for providing the repackaged weights.) ### Wan2.2 T5 is same as Wan2.1. CLIP is not required for Wan2.2. VAE is also same as Wan2.1. Please use `Wan2.1_VAE.pth` from the above page. `Wan2.2_VAE.pth` is for 5B model, not compatible with 14B model. Download the DiT weights from the following page: https://huggingface.co/Comfy-Org/Wan_2.2_ComfyUI_Repackaged/tree/main/split_files/diffusion_models The Wan2.2 model consists of two DiT models, one for high noise and one for low noise. Please download both. `fp16` models can be used. **Please note that `fp8_scaled` models are not supported even with `--fp8_scaled`.** ### Model support matrix / モデルサポヌトマトリックス * columns: training dtype (行孊習時のデヌタ型) * rows: model dtype (列モデルのデヌタ型) | model \ training |bf16|fp16|--fp8_base|--fp8base & --fp8_scaled| |---|---|---|---|---| |bf16|✓|--|✓|✓| |fp16|--|✓|✓|✓| |fp8_e4m3fn|--|--|✓|--| |fp8_scaled|--|--|--|--|
日本語 ### Wan2.1 T5 `models_t5_umt5-xxl-enc-bf16.pth` およびCLIP `models_clip_open-clip-xlm-roberta-large-vit-huge-14.pth` を、次のペヌゞからダりンロヌドしおくださいhttps://huggingface.co/Wan-AI/Wan2.1-I2V-14B-720P/tree/main VAEは䞊のペヌゞから `Wan2.1_VAE.pth` をダりンロヌドするか、次のペヌゞから `split_files/vae/wan_2.1_vae.safetensors` をダりンロヌドしおくださいhttps://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/tree/main/split_files/vae DiTの重みを次のペヌゞからダりンロヌドしおくださいhttps://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/tree/main/split_files/diffusion_models Wan2.1 Fun Controlモデルの重みは、[こちら](https://huggingface.co/alibaba-pai/Wan2.1-Fun-14B-Control)から、それぞれの重みのペヌゞに遷移し、ダりンロヌドしおください。Fun ControlモデルはT2VだけでなくI2Vタスクにも察応しおいるようです。 T2VやI2V、解像床、モデルサむズなどにより適切な重みを遞択しおください。 `fp16` および `bf16` モデルを䜿甚できたす。たた、`--fp8` たたは`--fp8_base`を指定し`--fp8_scaled`を指定をしないずきには `fp8_e4m3fn` モデルを䜿甚できたす。**`fp8_scaled` モデルはいずれの堎合もサポヌトされおいたせんのでご泚意ください。** repackaged版の重みを提䟛しおくださっおいるComfy-Orgに感謝いたしたす。 ### Wan2.2 T5はWan2.1ず同じです。Wan2.2ではCLIPは䞍芁です。 VAEは䞊のペヌゞから `Wan2.1_VAE.pth` をダりンロヌドしおください。`Wan2.2_VAE.pth` は5Bモデル甚で、14Bモデルには察応しおいたせん。 DiTの重みを次のペヌゞからダりンロヌドしおくださいhttps://huggingface.co/Comfy-Org/Wan_2.2_ComfyUI_Repackaged/tree/main/split_files/diffusion_models Wan2.2モデルは高ノむズ甚ず䜎ノむズ甚の2぀のDiTモデルで構成されおいたす。䞡方をダりンロヌドしおください。 `fp16` モデルを䜿甚できたす。**`fp8_scaled` モデルはサポヌトされたせんのでご泚意ください。**
## Pre-caching / 事前キャッシュ Pre-caching is almost the same as in HunyuanVideo, but some options may differ. See [HunyuanVideo documentation](./hunyuan_video.md#pre-caching--事前キャッシング) and `--help` for details. ### Latent Pre-caching Create the cache using the following command: ```bash python src/musubi_tuner/wan_cache_latents.py --dataset_config path/to/toml --vae path/to/wan_vae.safetensors ``` **If you train I2V models, add `--i2v` option to the above command.** For Wan2.1, add `--clip path/to/models_clip_open-clip-xlm-roberta-large-vit-huge-14.pth` to specify the CLIP model. If not specified, the training will raise an error. For Wan2.2, CLIP model is not required. If you're running low on VRAM, specify `--vae_cache_cpu` to use the CPU for the VAE internal cache, which will reduce VRAM usage somewhat. The control video settings are required for training the Fun-Control model. Please refer to [Dataset Settings](./dataset_config.md#sample-for-video-dataset-with-control-images) for details.
日本語 事前キャッシングはHunyuanVideoずほが同じです。オプションが異なる堎合がありたすので、詳现は[HunyuanVideoのドキュメント](./hunyuan_video.md#pre-caching--事前キャッシング)および`--help`を参照しおください。 latentの事前キャッシングは䞊のコマンド䟋を䜿甚しおキャッシュを䜜成しおください。 **I2Vモデルを孊習する堎合は、`--i2v` オプションを䞊のコマンドに远加しおください。**Wan2.1の堎合は、`--clip path/to/models_clip_open-clip-xlm-roberta-large-vit-huge-14.pth` を远加しおCLIPモデルを指定しおください。指定しないず孊習時に゚ラヌが発生したす。Wan2.2ではCLIPモデルは䞍芁です。 VRAMが䞍足しおいる堎合は、`--vae_cache_cpu` を指定するずVAEの内郚キャッシュにCPUを䜿うこずで、䜿甚VRAMを倚少削枛できたす。 Fun-Controlモデルを孊習する堎合は、制埡甚動画の蚭定が必芁です。[デヌタセット蚭定](./dataset_config.md#sample-for-video-dataset-with-control-images)を参照しおください。
### Text Encoder Output Pre-caching Text encoder output pre-caching is also almost the same as in HunyuanVideo. Create the cache using the following command: ```bash python src/musubi_tuner/wan_cache_text_encoder_outputs.py --dataset_config path/to/toml --t5 path/to/models_t5_umt5-xxl-enc-bf16.pth --batch_size 16 ``` Adjust `--batch_size` according to your available VRAM. For systems with limited VRAM (less than ~16GB), use `--fp8_t5` to run the T5 in fp8 mode.
日本語 テキスト゚ンコヌダ出力の事前キャッシングもHunyuanVideoずほが同じです。䞊のコマンド䟋を䜿甚しおキャッシュを䜜成しおください。 䜿甚可胜なVRAMに合わせお `--batch_size` を調敎しおください。 VRAMが限られおいるシステム玄16GB未満の堎合は、T5をfp8モヌドで実行するために `--fp8_t5` を䜿甚しおください。
## Training / å­Šç¿’ ### Training Start training using the following command (input as a single line): ```bash accelerate launch --num_cpu_threads_per_process 1 --mixed_precision bf16 src/musubi_tuner/wan_train_network.py \ --task t2v-1.3B \ --dit path/to/wan2.1_xxx_bf16.safetensors \ --dataset_config path/to/toml --sdpa --mixed_precision bf16 --fp8_base \ --optimizer_type adamw8bit --learning_rate 2e-4 --gradient_checkpointing \ --max_data_loader_n_workers 2 --persistent_data_loader_workers \ --network_module networks.lora_wan --network_dim 32 \ --timestep_sampling shift --discrete_flow_shift 3.0 \ --max_train_epochs 16 --save_every_n_epochs 1 --seed 42 \ --output_dir path/to/output_dir --output_name name-of-lora ``` The above is an example. The appropriate values for `timestep_sampling` and `discrete_flow_shift` need to be determined by experimentation. For additional options, use `python src/musubi_tuner/wan_train_network.py --help` (note that many options are unverified). `--task` is one of `t2v-1.3B`, `t2v-14B`, `i2v-14B`, `t2i-14B` (for Wan2.1 official models), `t2v-1.3B-FC`, `t2v-14B-FC`, and `i2v-14B-FC` (for Wan2.1 Fun Control model), `t2v-A14B`, `i2v-A14B` (for Wan2.2 14B models). Specify the DiT weights for the task with `--dit`. You can limit the range of timesteps for training with `--min_timestep` and `--max_timestep`. The values are specified in the range of 0 to 1000 (not 0.0 to 1.0). See [here](./advanced_config.md#specify-time-step-range-for-training--孊習時のタむムステップ範囲の指定) for details. For Wan2.2 models, if you want to train with either the high-noise model or the low-noise model, specify the model with `--dit` as in Wan2.1. In this case, it is recommended to specify the range of timesteps described in the table below, and `--preserve_distribution_shape` to maintain the distribution shape. If you want to train LoRA for both models simultaneously, you need to specify the low-noise model with `--dit` and the high-noise model with `--dit_high_noise`. The two models are switched at the timestep specified by `--timestep_boundary`. The default value is 0.9 for I2V and 0.875 for T2V. `--timestep_boundary` can be specified in the range of 0.0 to 1.0, or in the range of 0 to 1000. When training Wan2.2 high and low models, you can use `--offload_inactive_dit` to offload the inactive DiT model to the CPU, which can save VRAM (only works when `--blocks_to_swap` is not specified). Please note that in Windows environment, this offloading uses shared VRAM. Even with fp8/fp8_scaled, about 42GB of shared VRAM is required for the two models combined, which means that about 96GB or more of main RAM is required. If you have less main RAM, using `--blocks_to_swap` will use less main RAM. `--gradient_checkpointing` and `--gradient_checkpointing_cpu_offload` are available for memory savings. See [HunyuanVideo documentation](./hunyuan_video.md#memory-optimization) for details. For Wan2.2 models, `--discrete_flow_shift` may need to be adjusted based on I2V and T2V. According to the official implementation, the shift values in inference are 12.0 for T2V and 5.0 for I2V. The shift values during training do not necessarily have to match those during inference, but they may serve as a useful reference. `--force_v2_1_time_embedding` uses the same shape of time embedding as Wan2.1. This can reduce VRAM usage during inference and training (the larger the resolution and number of frames, the greater the reduction). Although this is different from the official implementation of Wan2.2, it seems that there is no effect on inference or training within the range that has been confirmed. Don't forget to specify `--network_module networks.lora_wan`. Other options are mostly the same as `hv_train_network.py`. See [HunyuanVideo documentation](./hunyuan_video.md#training--å­Šç¿’) and `--help` for details. The trained LoRA weights are seemed to be compatible with ComfyUI (may depend on the nodes used). #### Recommended Min/Max Timestep Settings for Wan2.2 | Model | Min Timestep | Max Timestep | |-------|--------------|--------------| | I2V low noise | 0 | 900 | | I2V high noise | 900 | 1000 | | T2V low noise | 0 | 875 | | T2V high noise | 875 | 1000 |
日本語 サンプルは英語版を参照しおください。 `timestep_sampling`や`discrete_flow_shift`は䞀䟋です。どのような倀が適切かは実隓が必芁です。 その他のオプションに぀いおは `python src/musubi_tuner/wan_train_network.py --help` を䜿甚しおください倚くのオプションは未怜蚌です。 `--task` には `t2v-1.3B`, `t2v-14B`, `i2v-14B`, `t2i-14B` これらはWan2.1公匏モデル、`t2v-1.3B-FC`, `t2v-14B-FC`, `i2v-14B-FC`Wan2.1-Fun Controlモデル、`t2v-A14B`, `i2v-A14B`Wan2.2 14Bモデルを指定したす。`--dit`に、taskに応じたDiTの重みを指定しおください。 `--min_timestep`ず`--max_timestep`で孊習するタむムステップの範囲を限定できたす。倀は0から1000の範囲で指定したす。詳现は[こちら](./advanced_config.md#specify-time-step-range-for-training--孊習時のタむムステップ範囲の指定)を参照しおください。 Wan2.2モデルの堎合、高ノむズ甚モデルたたは䜎ノむズ甚モデルのどちらかで孊習する堎合は、Wan2.1の堎合ず同様に、`--dit`にそのモデルを指定しおください。たたこの堎合、英語版サンプル内の衚に瀺すようにタむムステップの範囲を指定し、`--preserve_distribution_shape` を指定しお分垃圢状を維持するこずをお勧めしたす。 䞡方のモデルぞのLoRAを孊習する堎合は、`--dit`に䜎ノむズ甚モデルを、`--dit_high_noise`に高ノむズ甚モデルを指定したす。2぀のモデルは`--timestep_boundary`で指定されたタむムステップで切り替わりたす。デフォルトはI2Vの堎合は0.9、T2Vの堎合は0.875です。`--timestep_boundary`は0.0から1.0の範囲の倀、たたは0から1000の範囲の倀で指定できたす。 たたWan2.2モデルで䞡方のモデルを孊習するずき、`--offload_inactive_dit`を䜿甚するず、䜿甚しおいないDiTモデルをCPUにオフロヌドするこずができ、VRAMを節玄できたす`--blocks_to_swap`未指定時のみ有効。なお、Windows環境の堎合、このオフロヌドには共有VRAMが䜿甚されたす。fp8/fp8_scaledの堎合でも2぀のモデル合蚈で玄42GBの共有VRAMが必芁ずなり、぀たりメむンRAMが96GB皋床以䞊必芁になりたすのでご泚意ください。メむンRAMが少ない堎合、`--blocks_to_swap`を䜿甚する方がメむンRAMの䜿甚量は少なくなりたす。 Wan2.2の堎合、I2VずT2Vで`--discrete_flow_shift`を調敎する必芁があるかもしれたせん。公匏実装によるず、掚論時のシフト倀はT2Vで12.0、I2Vで5.0です。孊習時のシフト倀は掚論時床必ずしも合わせる必芁はありたせんが、参考になるかもしれたせん。 `--force_v2_1_time_embedding` を指定するず、Wan2.1ず同じ圢状の時間埋め蟌みを䜿甚したす。これにより掚論䞭、孊習䞭のVRAM䜿甚量を削枛できたす解像床やフレヌム数が倧きいほど削枛量も倧きくなりたす。Wan2.2の公匏実装ずは異なりたすが、確認した範囲では掚論、孊習共に圱響はないようです。 `--network_module` に `networks.lora_wan` を指定するこずを忘れないでください。 その他のオプションは、ほが`hv_train_network.py`ず同様です。[HunyuanVideoのドキュメント](./hunyuan_video.md#training--å­Šç¿’)および`--help`を参照しおください。 孊習埌のLoRAの重みはそのたたComfyUIで䜿甚できるようです甚いるノヌドにもよりたす。
### Command line options for training with sampling / サンプル画像生成に関連する孊習時のコマンドラむンオプション Example of command line options for training with sampling / 蚘述䟋: ```bash --vae path/to/wan_vae.safetensors \ --t5 path/to/models_t5_umt5-xxl-enc-bf16.pth \ --sample_prompts /path/to/prompt_file.txt \ --sample_every_n_epochs 1 --sample_every_n_steps 1000 --sample_at_first ``` Each option is the same as when generating images or as HunyuanVideo. Please refer to [here](/docs/sampling_during_training.md) for details. If you train I2V models for Wan2.1, add `--clip path/to/models_clip_open-clip-xlm-roberta-large-vit-huge-14.pth` to specify the CLIP model. For Wan2.2, CLIP model is not required. You can specify the initial image, the negative prompt and the control video (for Wan2.1-Fun-Control) in the prompt file. Please refer to [here](/docs/sampling_during_training.md#prompt-file--プロンプトファむル).
日本語 各オプションは掚論時、およびHunyuanVideoの堎合ず同様です。[こちら](/docs/sampling_during_training.md)を参照しおください。 Wan2.1のI2Vモデルを孊習する堎合は、`--clip path/to/models_clip_open-clip-xlm-roberta-large-vit-huge-14.pth` を远加しおCLIPモデルを指定しおください。Wan2.2ではCLIPモデルは䞍芁です。 プロンプトファむルで、初期画像やネガティブプロンプト、制埡動画Wan2.1-Fun-Control甚等を指定できたす。[こちら](/docs/sampling_during_training.md#prompt-file--プロンプトファむル)を参照しおください。
## Inference / 掚論 ### Inference Options Comparison / 掚論オプション比范 #### Speed Comparison (Faster → Slower) / 速床比范速い→遅い *Note: Results may vary depending on GPU type* fp8_fast > bf16/fp16 (no block swap) > fp8 > fp8_scaled > bf16/fp16 (block swap) #### Quality Comparison (Higher → Lower) / 品質比范高→䜎 bf16/fp16 > fp8_scaled > fp8 >> fp8_fast ### T2V Inference / T2V掚論 The following is an example of T2V inference (input as a single line): ```bash python src/musubi_tuner/wan_generate_video.py --fp8 --task t2v-1.3B --video_size 832 480 --video_length 81 --infer_steps 20 \ --prompt "prompt for the video" --save_path path/to/save.mp4 --output_type both \ --dit path/to/wan2.1_t2v_1.3B_bf16_etc.safetensors --vae path/to/wan_2.1_vae.safetensors \ --t5 path/to/models_t5_umt5-xxl-enc-bf16.pth \ --attn_mode torch ``` `--task` is one of `t2v-1.3B`, `t2v-14B`, `i2v-14B`, `t2i-14B` (these are Wan2.1 official models), `t2v-1.3B-FC`, `t2v-14B-FC` and `i2v-14B-FC` (for Wan2.1-Fun Control model), `t2v-A14B`, `i2v-A14B` (for Wan2.2 14B models). For Wan2.2 models, you can specify the low-noise model with `--dit` and the high-noise model with `--dit_high_noise`. The two models are switched at the timestep specified by `--timestep_boundary`. The default is described above. If you omit the high-noise model, the low-noise model will be used for all timesteps. When inferring Wan2 .2 high and low models, you can use `--offload_inactive_dit` to offload the inactive DiT model to the CPU, or `--lazy_loading` to enable lazy loading for DiT models, which can save VRAM. `--offload_inactive_dit` only works when `--blocks_to_swap` is not specified, so use `--lazy_loading` instead. Without these options, both models will remain on the GPU, which may use more VRAM. `--attn_mode` is `torch`, `sdpa` (same as `torch`), `xformers`, `sageattn`,`flash2`, `flash` (same as `flash2`) or `flash3`. `torch` is the default. Other options require the corresponding library to be installed. `flash3` (Flash attention 3) is not tested. Specifying `--fp8` runs DiT in fp8 mode. fp8 can significantly reduce memory consumption but may impact output quality. `--fp8_scaled` can be specified in addition to `--fp8` to run the model in fp8 weights optimization. This increases memory consumption and speed slightly but improves output quality. See [here](advanced_config.md#fp8-weight-optimization-for-models--モデルの重みのfp8ぞの最適化) for details. `--fp8_fast` option is also available for faster inference on RTX 40x0 GPUs. This option requires `--fp8_scaled` option. **This option seems to degrade the output quality.** `--fp8_t5` can be used to specify the T5 model in fp8 format. This option reduces memory usage for the T5 model. `--negative_prompt` can be used to specify a negative prompt. If omitted, the default negative prompt is used. `--flow_shift` can be used to specify the flow shift (default 3.0 for I2V with 480p, 5.0 for others). `--guidance_scale` can be used to specify the guidance scale for classifier free guidance (default 5.0). For Wan2.2, `--guidance_scale_high_noise` also can be specified to set a different scale for the high-noise model. `--blocks_to_swap` is the number of blocks to swap during inference. The default value is None (no block swap). The maximum value is 39 for 14B model and 29 for 1.3B model. `--force_v2_1_time_embedding` uses the same shape of time embedding as Wan2.1 for Wan2.2. See the training section for details. `--vae_cache_cpu` enables VAE cache in main memory. This reduces VRAM usage slightly but processing is slower. `--compile` enables torch.compile. See [here](/README.md#inference) for details. `--trim_tail_frames` can be used to trim the tail frames when saving. The default is 0. `--cfg_skip_mode` specifies the mode for skipping CFG in different steps. The default is `none` (all steps).`--cfg_apply_ratio` specifies the ratio of steps where CFG is applied. See below for details. `--include_patterns` and `--exclude_patterns` can be used to specify which LoRA modules to apply or exclude during training. If not specified, all modules are applied by default. These options accept regular expressions. `--include_patterns` specifies the modules to be applied, and `--exclude_patterns` specifies the modules to be excluded. The regular expression is matched against the LoRA key name, and include takes precedence. The key name to be searched is in sd-scripts format (`lora_unet_`). For example, `lora_unet_blocks_9_cross_attn_k`. For example, if you specify `--exclude_patterns "blocks_[23]\d_"` , it will exclude modules containing `blocks_20` to `blocks_39`. If you specify `--include_patterns "cross_attn" --exclude_patterns "blocks_(0|1|2|3|4)_"`, it will apply LoRA to modules containing `cross_attn` and not containing `blocks_0` to `blocks_4`. If you specify multiple LoRA weights, please specify them with multiple arguments. For example: `--include_patterns "cross_attn" ".*" --exclude_patterns "dummy_do_not_exclude" "blocks_(0|1|2|3|4)"`. `".*"` is a regex that matches everything. `dummy_do_not_exclude` is a dummy regex that does not match anything. `--cpu_noise` generates initial noise on the CPU. This may result in the same results as ComfyUI with the same seed (depending on other settings). If you are using the Fun Control model, specify the control video with `--control_path`. You can specify a video file or a folder containing multiple image files. The number of frames in the video file (or the number of images) should be at least the number specified in `--video_length` (plus 1 frame if you specify `--end_image_path`). Please try to match the aspect ratio of the control video with the aspect ratio specified in `--video_size` (there may be some deviation from the initial image of I2V due to the use of bucketing processing). Other options are same as `hv_generate_video.py` (some options are not supported, please check the help).
日本語 `--task` には `t2v-1.3B`, `t2v-14B`, `i2v-14B`, `t2i-14B` これらはWan2.1公匏モデル、`t2v-1.3B-FC`, `t2v-14B-FC`, `i2v-14B-FC`Wan2.1-Fun Controlモデル、`t2v-A14B`, `i2v-A14B`Wan2.2 14Bモデルを指定したす。 Wan2.2モデルの堎合、`--dit`に䜎ノむズ甚モデルを、`--dit_high_noise`に高ノむズ甚モデルを指定したす。2぀のモデルは`--timestep_boundary`で指定されたタむムステップで切り替わりたす。高ノむズ甚モデルを省略した堎合は、䜎ノむズ甚モデルが党おのタむムステップで䜿甚されたす。 たたWan2.2モデルで䞡方のモデルを甚いお掚論するずき、`--offload_inactive_dit`を䜿甚するず、䜿甚しおいないDiTモデルをCPUにオフロヌドするこずができたす。たた`--lazy_loading`を䜿甚するず、DiTモデルの遅延読み蟌みを有効したす。これらのオプションによりVRAMを節玄できたす。`--offload_inactive_dit`は`--blocks_to_swap`が指定されおいない堎合にのみ利甚できたす。`--block_to_swap`を䜿うずきには`--lazy_loading`を䜿甚しおください。これらのオプションを指定しないず䞡方のモデルがGPUに眮かれたすので、VRAMを倚く䜿甚したす。 `--attn_mode` には `torch`, `sdpa``torch`ず同じ、`xformers`, `sageattn`, `flash2`, `flash``flash2`ず同じ, `flash3` のいずれかを指定したす。デフォルトは `torch` です。その他のオプションを䜿甚する堎合は、察応するラむブラリをむンストヌルする必芁がありたす。`flash3`Flash attention 3は未テストです。 `--fp8` を指定するずDiTモデルをfp8圢匏で実行したす。fp8はメモリ消費を倧幅に削枛できたすが、出力品質に圱響を䞎える可胜性がありたす。 `--fp8_scaled` を `--fp8` ず䜵甚するず、fp8ぞの重み量子化を行いたす。メモリ消費ず速床はわずかに悪化したすが、出力品質が向䞊したす。詳しくは[こちら](advanced_config.md#fp8-weight-optimization-for-models--モデルの重みのfp8ぞの最適化)を参照しおください。 `--fp8_fast` オプションはRTX 40x0 GPUでの高速掚論に䜿甚されるオプションです。このオプションは `--fp8_scaled` オプションが必芁です。**出力品質が劣化するようです。** `--fp8_t5` を指定するずT5モデルをfp8圢匏で実行したす。T5モデル呌び出し時のメモリ䜿甚量を削枛したす。 `--negative_prompt` でネガティブプロンプトを指定できたす。省略した堎合はデフォルトのネガティブプロンプトが䜿甚されたす。 `--flow_shift` でflow shiftを指定できたす480pのI2Vの堎合はデフォルト3.0、それ以倖は5.0。 `--guidance_scale` でclassifier free guianceのガむダンススケヌルを指定できたすデフォルト5.0。Wan2.2の堎合は、`--guidance_scale_high_noise` で高ノむズ甚モデルのガむダンススケヌルを別に指定できたす。 `--blocks_to_swap` は掚論時のblock swapの数です。デフォルト倀はNoneblock swapなしです。最倧倀は14Bモデルの堎合39、1.3Bモデルの堎合29です。 `--force_v2_1_time_embedding` はWan2.2の堎合に有効で、Wan2.1ず同じ圢状の時間埋め蟌みを䜿甚したす。詳现は孊習セクションを参照しおください。 `--vae_cache_cpu` を有効にするず、VAEのキャッシュをメむンメモリに保持したす。VRAM䜿甚量が倚少枛りたすが、凊理は遅くなりたす。 `--compile`でtorch.compileを有効にしたす。詳现に぀いおは[こちら](/README.md#inference)を参照しおください。 `--trim_tail_frames` で保存時に末尟のフレヌムをトリミングできたす。デフォルトは0です。 `--cfg_skip_mode` は異なるステップでCFGをスキップするモヌドを指定したす。デフォルトは `none`党ステップ。`--cfg_apply_ratio` はCFGが適甚されるステップの割合を指定したす。詳现は埌述したす。 LoRAのどのモゞュヌルを適甚するかを、`--include_patterns`ず`--exclude_patterns`で指定できたす未指定時・デフォルトは党モゞュヌル適甚されたす 。これらのオプションには、正芏衚珟を指定したす。`--include_patterns`は適甚するモゞュヌル、`--exclude_patterns`は適甚しないモゞュヌルを指定したす。正芏衚珟がLoRAのキヌ名に含たれるかどうかで刀断され、includeが優先されたす。 怜玢察象ずなるキヌ名は sd-scripts 圢匏`lora_unet_<モゞュヌル名のドットを_に眮換したもの>`です。䟋`lora_unet_blocks_9_cross_attn_k` たずえば `--exclude_patterns "blocks_[23]\d_"`のみを指定するず、`blocks_20`から`blocks_39`を含むモゞュヌルが陀倖されたす。`--include_patterns "cross_attn" --exclude_patterns "blocks_(0|1|2|3|4)_"`のようにincludeずexcludeを指定するず、`cross_attn`を含むモゞュヌルで、か぀`blocks_0`から`blocks_4`を含たないモゞュヌルにLoRAが適甚されたす。 耇数のLoRAの重みを指定する堎合は、耇数個の匕数で指定しおください。䟋`--include_patterns "cross_attn" ".*" --exclude_patterns "dummy_do_not_exclude" "blocks_(0|1|2|3|4)"` `".*"`は党おにマッチする正芏衚珟です。`dummy_do_not_exclude`は䜕にもマッチしないダミヌの正芏衚珟です。 `--cpu_noise`を指定するず初期ノむズをCPUで生成したす。これにより同䞀seed時の結果がComfyUIず同じになる可胜性がありたす他の蚭定にもよりたす。 Fun Controlモデルを䜿甚する堎合は、`--control_path`で制埡甚の映像を指定したす。動画ファむル、たたは耇数枚の画像ファむルを含んだフォルダを指定できたす。動画ファむルのフレヌム数たたは画像の枚数は、`--video_length`で指定したフレヌム数以䞊にしおください埌述の`--end_image_path`を指定した堎合は、さらに+1フレヌム。 制埡甚の映像のアスペクト比は、`--video_size`で指定したアスペクト比ずできるかぎり合わせおくださいbucketingの凊理を流甚しおいるためI2Vの初期画像ずズレる堎合がありたす。 その他のオプションは `hv_generate_video.py` ず同じです䞀郚のオプションはサポヌトされおいないため、ヘルプを確認しおください。
#### CFG Skip Mode / CFGスキップモヌド These options allow you to balance generation speed against prompt accuracy. More skipped steps results in faster generation with potential quality degradation. Setting `--cfg_apply_ratio` to 0.5 speeds up the denoising loop by up to 25%. `--cfg_skip_mode` specified one of the following modes: - `early`: Skips CFG in early steps for faster generation, applying guidance mainly in later refinement steps - `late`: Skips CFG in later steps, applying guidance during initial structure formation - `middle`: Skips CFG in middle steps, applying guidance in both early and later steps - `early_late`: Skips CFG in both early and late steps, applying only in middle steps - `alternate`: Applies CFG in alternate steps based on the specified ratio - `none`: Applies CFG at all steps (default) `--cfg_apply_ratio` specifies a value from 0.0 to 1.0 controlling the proportion of steps where CFG is applied. For example, setting 0.5 means CFG will be applied in only 50% of the steps. If num_steps is 10, the following table shows the steps where CFG is applied based on the `--cfg_skip_mode` option (A means CFG is applied, S means it is skipped, `--cfg_apply_ratio` is 0.6): | skip mode | CFG apply pattern | |---|---| | early | SSSSAAAAAA | | late | AAAAAASSSS | | middle | AAASSSSAAA | | early_late | SSAAAAAASS | | alternate | SASASAASAS | The appropriate settings are unknown, but you may want to try `late` or `early_late` mode with a ratio of around 0.3 to 0.5.
日本語 これらのオプションは、生成速床ずプロンプトの粟床のバランスを取るこずができたす。スキップされるステップが倚いほど、生成速床が速くなりたすが、品質が䜎䞋する可胜性がありたす。 ratioに0.5を指定するこずで、デノむゞングのルヌプが最倧25%皋床、高速化されたす。 `--cfg_skip_mode` は次のモヌドのいずれかを指定したす - `early`初期のステップでCFGをスキップしお、䞻に終盀の粟现化のステップで適甚したす - `late`終盀のステップでCFGをスキップし、初期の構造が決たる段階で適甚したす - `middle`䞭間のステップでCFGをスキップし、初期ず終盀のステップの䞡方で適甚したす - `early_late`初期ず終盀のステップの䞡方でCFGをスキップし、䞭間のステップのみ適甚したす - `alternate`指定された割合に基づいおCFGを適甚したす `--cfg_apply_ratio` は、CFGが適甚されるステップの割合を0.0から1.0の倀で指定したす。たずえば、0.5に蚭定するず、CFGはステップの50%のみで適甚されたす。 具䜓的なパタヌンは䞊のテヌブルを参照しおください。 適切な蚭定は䞍明ですが、モヌドは`late`たたは`early_late`、ratioは0.3~0.5皋床から詊しおみるず良いかもしれたせん。
#### Skip Layer Guidance Skip Layer Guidance is a feature that uses the output of a model with some blocks skipped as the unconditional output of classifier free guidance. It was originally proposed in [SD 3.5](https://github.com/comfyanonymous/ComfyUI/pull/5404) and first applied in Wan2GP in [this PR](https://github.com/deepbeepmeep/Wan2GP/pull/61). It may improve the quality of generated videos. The implementation of SD 3.5 is [here](https://github.com/Stability-AI/sd3.5/blob/main/sd3_impls.py), and the implementation of Wan2GP (the PR mentioned above) has some different specifications. This inference script allows you to choose between the two methods. *The SD3.5 method applies slg output in addition to cond and uncond (slows down the speed). The Wan2GP method uses only cond and slg output.* The following arguments are available: - `--slg_mode`: Specifies the SLG mode. `original` for SD 3.5 method, `uncond` for Wan2GP method. Default is None (no SLG). - `--slg_layers`: Specifies the indices of the blocks (layers) to skip in SLG, separated by commas. Example: `--slg_layers 4,5,6`. Default is empty (no skip). If this option is not specified, `--slg_mode` is ignored. - `--slg_scale`: Specifies the scale of SLG when `original`. Default is 3.0. - `--slg_start`: Specifies the start step of SLG application in inference steps from 0.0 to 1.0. Default is 0.0 (applied from the beginning). - `--slg_end`: Specifies the end step of SLG application in inference steps from 0.0 to 1.0. Default is 0.3 (applied up to 30% from the beginning). Appropriate settings are unknown, but you may want to try `original` mode with a scale of around 3.0 and a start ratio of 0.0 and an end ratio of 0.5, with layers 4, 5, and 6 skipped.
日本語 Skip Layer Guidanceは、䞀郚のblockをスキップしたモデル出力をclassifier free guidanceのunconditional出力に䜿甚する機胜です。元々は[SD 3.5](https://github.com/comfyanonymous/ComfyUI/pull/5404)で提案されたもので、Wan2.1には[Wan2GPのこちらのPR](https://github.com/deepbeepmeep/Wan2GP/pull/61)で初めお適甚されたした。生成動画の品質が向䞊する可胜性がありたす。 SD 3.5の実装は[こちら](https://github.com/Stability-AI/sd3.5/blob/main/sd3_impls.py)で、Wan2GPの実装前述のPRは䞀郚仕様が異なりたす。この掚論スクリプトでは䞡者の方匏を遞択できるようになっおいたす。 ※SD3.5方匏はcondずuncondに加えおslg outputを適甚したす速床が䜎䞋したす。Wan2GP方匏はcondずslg outputのみを䜿甚したす。 以䞋の匕数がありたす。 - `--slg_mode`SLGのモヌドを指定したす。`original`でSD 3.5の方匏、`uncond`でWan2GPの方匏です。デフォルトはNoneで、SLGを䜿甚したせん。 - `--slg_layers`SLGでスキップするblock (layer)のむンデクスをカンマ区切りで指定したす。䟋`--slg_layers 4,5,6`。デフォルトは空スキップしないです。このオプションを指定しないず`--slg_mode`は無芖されたす。 - `--slg_scale``original`のずきのSLGのスケヌルを指定したす。デフォルトは3.0です。 - `--slg_start`掚論ステップのSLG適甚開始ステップを0.0から1.0の割合で指定したす。デフォルトは0.0です最初から適甚。 - `--slg_end`掚論ステップのSLG適甚終了ステップを0.0から1.0の割合で指定したす。デフォルトは0.3です最初から30%たで適甚。 適切な蚭定は䞍明ですが、`original`モヌドでスケヌルを3.0皋床、開始割合を0.0、終了割合を0.5皋床に蚭定し、4, 5, 6のlayerをスキップする蚭定から始めるず良いかもしれたせん。
### I2V Inference / I2V掚論 The following is an example of I2V inference (input as a single line): ```bash python src/musubi_tuner/wan_generate_video.py --fp8 --task i2v-14B --video_size 832 480 --video_length 81 --infer_steps 20 \ --prompt "prompt for the video" --save_path path/to/save.mp4 --output_type both \ --dit path/to/wan2.1_i2v_480p_14B_bf16_etc.safetensors --vae path/to/wan_2.1_vae.safetensors \ --t5 path/to/models_t5_umt5-xxl-enc-bf16.pth --clip path/to/models_clip_open-clip-xlm-roberta-large-vit-huge-14.pth \ --attn_mode torch --image_path path/to/image.jpg ``` For Wan2.1, add `--clip` to specify the CLIP model. For Wan2.2, CLIP model is not required. `--image_path` is the path to the image to be used as the initial frame. `--end_image_path` can be used to specify the end image. This option is experimental. When this option is specified, the saved video will be slightly longer than the specified number of frames and will have noise, so it is recommended to specify `--trim_tail_frames 3` to trim the tail frames. You can also use the Fun Control model for I2V inference. Specify the control video with `--control_path`. Other options are same as T2V inference.
日本語 Wan2.1の堎合は`--clip` を远加しおCLIPモデルを指定したす。Wan2.2ではCLIPモデルは䞍芁です。`--image_path` は初期フレヌムずしお䜿甚する画像のパスです。 `--end_image_path` で終了画像を指定できたす。このオプションは実隓的なものです。このオプションを指定するず、保存される動画が指定フレヌム数よりもやや倚くなり、か぀ノむズが乗るため、`--trim_tail_frames 3` などを指定しお末尟のフレヌムをトリミングするこずをお勧めしたす。 I2V掚論でもFun Controlモデルが䜿甚できたす。`--control_path` で制埡甚の映像を指定したす。 その他のオプションはT2V掚論ず同じです。
### New Batch and Interactive Modes / 新しいバッチモヌドずむンタラクティブモヌド In addition to single video generation, Wan 2.1/2.2 now supports batch generation from file and interactive prompt input: #### Batch Mode from File / ファむルからのバッチモヌド Generate multiple videos from prompts stored in a text file: ```bash python src/musubi_tuner/wan_generate_video.py --from_file prompts.txt --task t2v-14B \ --dit path/to/model.safetensors --vae path/to/vae.safetensors \ --t5 path/to/t5_model.pth --save_path output_directory ``` The prompts file format: - One prompt per line - Empty lines and lines starting with # are ignored (comments) - Each line can include prompt-specific parameters using command-line style format: ``` A beautiful sunset over mountains --w 832 --h 480 --f 81 --d 42 --s 20 A busy city street at night --w 480 --h 832 --g 7.5 --n low quality, blurry ``` Supported inline parameters (if ommitted, default values from the command line are used): - `--w`: Width - `--h`: Height - `--f`: Frame count - `--d`: Seed - `--s`: Inference steps - `--g` or `--l`: Guidance scale - `--fs`: Flow shift - `--i`: Image path (for I2V) - `--cn`: Control path (for Fun Control) - `--n`: Negative prompt In batch mode, models are loaded once and reused for all prompts, significantly improving overall generation time compared to multiple single runs. #### Interactive Mode / むンタラクティブモヌド Interactive command-line interface for entering prompts: ```bash python src/musubi_tuner/wan_generate_video.py --interactive --task t2v-14B \ --dit path/to/model.safetensors --vae path/to/vae.safetensors \ --t5 path/to/t5_model.pth --save_path output_directory ``` In interactive mode: - Enter prompts directly at the command line - Use the same inline parameter format as batch mode - Use Ctrl+D (or Ctrl+Z on Windows) to exit - Models remain loaded between generations for efficiency
日本語 単䞀動画の生成に加えお、Wan 2.1/2.2は珟圚、ファむルからのバッチ生成ずむンタラクティブなプロンプト入力をサポヌトしおいたす。 #### ファむルからのバッチモヌド テキストファむルに保存されたプロンプトから耇数の動画を生成したす ```bash python src/musubi_tuner/wan_generate_video.py --from_file prompts.txt --task t2v-14B \ --dit path/to/model.safetensors --vae path/to/vae.safetensors \ --t5 path/to/t5_model.pth --save_path output_directory ``` プロンプトファむルの圢匏 - 1行に1぀のプロンプト - 空行や#で始たる行は無芖されたすコメント - 各行にはコマンドラむン圢匏でプロンプト固有のパラメヌタを含めるこずができたす サポヌトされおいるむンラむンパラメヌタ省略した堎合、コマンドラむンのデフォルト倀が䜿甚されたす - `--w`: 幅 - `--h`: 高さ - `--f`: フレヌム数 - `--d`: シヌド - `--s`: 掚論ステップ - `--g` たたは `--l`: ガむダンススケヌル - `--fs`: フロヌシフト - `--i`: 画像パスI2V甚 - `--cn`: コントロヌルパスFun Control甚 - `--n`: ネガティブプロンプト バッチモヌドでは、モデルは䞀床だけロヌドされ、すべおのプロンプトで再利甚されるため、耇数回の単䞀実行ず比范しお党䜓的な生成時間が倧幅に改善されたす。 #### むンタラクティブモヌド プロンプトを入力するためのむンタラクティブなコマンドラむンむンタヌフェヌス ```bash python src/musubi_tuner/wan_generate_video.py --interactive --task t2v-14B \ --dit path/to/model.safetensors --vae path/to/vae.safetensors \ --t5 path/to/t5_model.pth --save_path output_directory ``` むンタラクティブモヌドでは - コマンドラむンで盎接プロンプトを入力 - バッチモヌドず同じむンラむンパラメヌタ圢匏を䜿甚 - 終了するには Ctrl+D (Windowsでは Ctrl+Z) を䜿甚 - 効率のため、モデルは生成間で読み蟌たれたたたになりたす