Wan 2.1 T2V 14B

'The full-size text-to-video model of Wan 2.1. A 14B-parameter DiT in the 8× spatial / 4× temporal latent space of the Wan 2.1 VAE, with UMT5-XXL text conditioning. Same architecture as each Wan 2.2 14B expert, as a single model.'

weights: Wan-AI/Wan2.1-T2V-14B-Diffusers ↗
org: Wan-AI (Alibaba)
modality: video
tasks: text-to-video · text-to-image
license: Apache 2.0
released: 2025-02-25not verified
native output: 1280×720 or 832×480, 81 frames, 16 fpsnot verified
total params: 20.10B
model.arch: wan21:14b

Components

role	model	params	size	dtype	trained
Transformer	Wan 2.1 T2V 14B DiT Wan-AI/Wan2.1-T2V-14B-Diffusers/transformer diffusers.WanTransformer3DModel Stored as fp32 in the Diffusers repo. By default ai-toolkit loads a ComfyUI repack instead (Comfy-Org/Wan_2.1_ComfyUI_repackaged: wan2.1_t2v_14B_fp8_scaled when quantizing with qfloat8, wan2.1_t2v_14B_bf16 otherwise), with the config from this folder.	14.29B 14,288,491,584	57.15 GB	fp32	yes
Text encoder	UMT5-XXL (encoder only) Wan-AI/Wan2.1-T2V-14B-Diffusers/text_encoder transformers.UMT5EncoderModel Stored as fp32 here (the Wan 2.2 repos ship it in bf16). Includes the 256k-token embedding table. ai-toolkit does not load this folder: it uses ai-toolkit/umt5_xxl_encoder (bf16, same parameter count), whose weights are swapped for the ComfyUI umt5_xxl file by default.	5.68B 5,680,910,336	22.72 GB	fp32	no
Tokenizer	UMT5 SentencePiece (256k vocab) Wan-AI/Wan2.1-T2V-14B-Diffusers/tokenizer transformers.T5TokenizerFast ai-toolkit loads the tokenizer from ai-toolkit/umt5_xxl_encoder instead.	—	—	—	—
VAE	Wan 2.1 VAE Wan-AI/Wan2.1-T2V-14B-Diffusers/vae diffusers.AutoencoderKLWan The same VAE is used by every Wan 2.1 model and by the Wan 2.2 14B models.	126.9M 126,892,531	508 MB	fp32	no
total		20.10B	80.39 GB

Latent space

spatial

8×

temporal

4×

channels

patch

1×2×2

autoencoder: Wan 2.1 VAE
pixels per token: 16×16 × 4 frames
frame count: 4n + 1
notes: Causal 3D VAE: 8× spatial downsampling and two temporal downsamples (temperal_downsample [false, true, true]). The first frame gets its own latent frame, which is why frame counts are 4n + 1. Latents are normalized with per-channel latents_mean / latents_std from the VAE config.

input	latent (c×t×h×w)	tokens
1024×1024 × 41f	16×11×128×128	45,056	ai-toolkit sample default
832×480 × 81f	16×21×60×104	32,760	native 480p
1280×720 × 81f	16×21×90×160	75,600	native 720p
1024×1024	16×1×128×128	4,096	still image

Architecture

Blocks: 40
Hidden size: 5120 (40 heads × 128)
FFN size: 13824
Text conditioning: Cross-attention on UMT5 hidden states (4096-d), 512 tokens max
Image conditioning: None (text-to-video only)
Timesteps: One timestep per sample
Objective: Rectified flow; the upstream scheduler uses shift 3.0
Norm / position: QK RMSNorm across heads, 3D RoPE

In AI Toolkit

model.arch: wan21:14b
UI label: Wan 2.1 (14B) (video)
model.name_or_path: Wan-AI/Wan2.1-T2V-14B-Diffusers
source: toolkit/models/wan21/wan21.py
toolkit/models/wan21/wan_lora_convert.py
toolkit/models/v2/diffusion_models/wan.py
toolkit/models/v2/text_encoders/umt5.py
toolkit/models/v2/vae/wan.py
toolkit/models/v2/resolver.py
extensions_built_in/diffusion_models/ui.tsx
extra UI sections: datasets.num_frames, model.low_vram, datasets.auto_frame_count

UI defaults

quantize / quantize_te: true / true (qfloat8)
noise_scheduler / sampler: flowmatch / flowmatch
timestep_type: sigmoid (UI default, not set by the model)
sample size: 1024×1024, 41 frames, 16 fps
datasets.fps: 16
network.conv: disabled (linear LoRA only)

Specifics

Arch string: The UI sends wan21:14b. ModelConfig drops everything after the colon, so it loads the same Wan21 class as wan21:1b; only name_or_path differs.
Resolution: Buckets snap to multiples of 16 (8× VAE × 2×2 patch).
Transformer source: For this repo id the weights come from Comfy-Org/Wan_2.1_ComfyUI_repackaged, with the config from name_or_path. quantize with qfloat8 picks wan2.1_t2v_14B_fp8_scaled; no quantization picks the bf16 file. A file already on disk wins over a download. A local folder loads as-is. model_kwargs.use_comfy_weights: false opts out.
Text encoder source: Loads ai-toolkit/umt5_xxl_encoder unless name_or_path is a local folder with a text_encoder/ subfolder. Its weights resolve to the ComfyUI umt5_xxl_fp8_e4m3fn_scaled file when quantize_te is on with qfloat8, otherwise umt5_xxl_fp16. Never trained.
VAE source: vae/ inside extras_name_or_path, which defaults to name_or_path.
Loss: Flow-matching target (noise − latents) over every latent frame. Training scheduler shift is 3.0.
LoRA target: WanTransformer3DModel
Saving: LoRAs are converted to the original Wan key layout, which ComfyUI can load. Full fine-tunes save a single file in the original layout.
Sampling: Uses flowmatch Euler with the training scheduler (shift 3.0). UniPC is disabled until a diffusers regression is fixed. low_vram also turns on VAE tiling and a pipeline that moves each model to the GPU only while it runs.
low_vram: Off in the UI defaults. The repo's 24 GB example config (train_lora_wan21_14b_24gb.yaml) turns it on with quantize: the transformer and text encoder load on the CPU and move to the GPU when needed.
Not supported: split_model_over_gpus, assistant_lora_path, inference_lora_path, lora_path
Metadata base version: wan_2.1

Example config

not verified

job: extensionconfig:  name: "my_wan21_14b_lora_v1"  process:    - type: "sd_trainer"      training_folder: "output"      device: cuda:0      network:        type: "lora"        linear: 32        linear_alpha: 32      save:        dtype: float16        save_every: 250        max_step_saves_to_keep: 4      datasets:        - folder_path: "/path/to/videos"          caption_ext: "txt"          caption_dropout_rate: 0.05          num_frames: 41          fps: 16          resolution: [480, 632]      train:        batch_size: 1        steps: 2000        gradient_checkpointing: true        noise_scheduler: "flowmatch"        timestep_type: "sigmoid"        optimizer: "adamw8bit"        lr: 1e-4        dtype: bf16        cache_text_embeddings: true      model:        name_or_path: "Wan-AI/Wan2.1-T2V-14B-Diffusers"        arch: "wan21:14b"        quantize: true        quantize_te: true        low_vram: true      sample:        sampler: "flowmatch"        sample_every: 250        width: 832        height: 480        num_frames: 41        fps: 16        guidance_scale: 5        sample_steps: 30        prompts:          - "woman playing the guitar, on stage, singing a song, laser lights, punk rocker"