Wan 2.1 I2V 14B 480P

'The 480p image-to-video model of Wan 2.1. A 14B DiT that takes the start frame twice: as a VAE latent concatenated onto the noisy input, and as CLIP ViT-H/14 features read through extra cross-attention. Uses the Wan 2.1 VAE and UMT5-XXL.'

weights: Wan-AI/Wan2.1-I2V-14B-480P-Diffusers ↗
org: Wan-AI (Alibaba)
modality: video
tasks: image-to-video
license: Apache 2.0
released: 2025-02-25not verified
native output: 832×480, 81 frames, 16 fpsnot verified
total params: 22.83B
model.arch: wan21_i2v:14b480p

Components

role	model	params	size	dtype	trained
Transformer	Wan 2.1 I2V 14B DiT (480P) Wan-AI/Wan2.1-I2V-14B-480P-Diffusers/transformer diffusers.WanTransformer3DModel 2,106,592,000 more parameters than the T2V 14B, mostly the image K/V projections added to every block. 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_i2v_480p_14B_fp8_scaled when quantizing with qfloat8, the bf16 file otherwise), with the config from this folder.	16.40B 16,395,083,584	65.58 GB	fp32	yes
Image encoder	CLIP ViT-H/14 vision tower (LAION-2B) Wan-AI/Wan2.1-I2V-14B-480P-Diffusers/image_encoder transformers.CLIPVisionModelWithProjection Config names laion/CLIP-ViT-H-14-laion2B-s32B-b79K. ai-toolkit loads it as CLIPVisionModel (no projection) and feeds the penultimate hidden states, 257 tokens × 1280, to the transformer.	632.1M 632,076,800	1.26 GB	bf16	no
Image processor	CLIP image processor (224×224) Wan-AI/Wan2.1-I2V-14B-480P-Diffusers/image_processor transformers.CLIPImageProcessor	—	—	—	—
Text encoder	UMT5-XXL (encoder only) Wan-AI/Wan2.1-I2V-14B-480P-Diffusers/text_encoder transformers.UMT5EncoderModel Stored as fp32 here. 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-I2V-14B-480P-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-I2V-14B-480P-Diffusers/vae diffusers.AutoencoderKLWan The same VAE as every other Wan 2.1 model and the Wan 2.2 14B models.	126.9M 126,892,531	508 MB	fp32	no
total		22.83B	90.08 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: The transformer takes 36 input channels: 16 noisy latent + 4 mask + 16 conditioning latent. The conditioning latent is the VAE encoding of the start frame followed by black frames for the rest of the clip. The mask is 1 for the start frame and 0 elsewhere; its 4 channels are the 4 pixel frames folded into each latent frame. Output is 16 channels.

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
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: Start-frame latent + mask concatenated on channels (in_channels 36), plus CLIP tokens (1280-d) through separate image K/V projections in each cross-attention
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_i2v:14b480p
UI label: Wan 2.1 I2V (14B-480P) (video)
model.name_or_path: Wan-AI/Wan2.1-I2V-14B-480P-Diffusers
source: toolkit/models/wan21/wan21_i2v.py
toolkit/models/wan21/wan21.py
toolkit/models/wan21/wan_utils.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/vision_encoders/clip_vision.py
toolkit/models/v2/resolver.py
extensions_built_in/diffusion_models/ui.tsx
extra UI sections: sample.ctrl_img, 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: weighted
sample size: 1024×1024, 41 frames, 16 fps
datasets.fps: 16
network.conv: disabled (linear LoRA only)

Specifics

Arch string: The UI sends wan21_i2v:14b480p. ModelConfig drops everything after the colon, so it loads the same Wan21I2V class as the 720P entry; only name_or_path differs.
Resolution: Buckets snap to multiples of 16 (8× VAE × 2×2 patch).
Image-to-video: Always on. Each step takes the first frame of the clip (or the still image), builds the 20-channel mask + conditioning latent by VAE-encoding that frame followed by black frames for the full clip length, and runs the frame through CLIP (bilinear resize to 224×224, no crop). Nothing is cached, so every step pays a full-length VAE encode.
Loss: Flow-matching target (noise − latents) over every latent frame, including the start frame. Training scheduler shift is 3.0.
Image encoder source: image_encoder/ and image_processor/ from extras_name_or_path (defaults to name_or_path), falling back to name_or_path. Never trained.
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_i2v_480p_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.
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: Every sample needs ctrl_img; sampling raises an error without one. Size is floored to a multiple of 16 and the control image is resized to it. Uses flowmatch Euler with the training scheduler (shift 3.0); UniPC is disabled until a diffusers regression is fixed.
low_vram: Keeps the text encoder and CLIP on the CPU between uses and swaps the VAE, CLIP, text encoder and transformer on and off the GPU while sampling.
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_i2v_480p_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: "weighted"        optimizer: "adamw8bit"        lr: 1e-4        dtype: bf16        cache_text_embeddings: true      model:        name_or_path: "Wan-AI/Wan2.1-I2V-14B-480P-Diffusers"        arch: "wan21_i2v:14b480p"        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        samples:          - prompt: "a woman turns and smiles at the camera"            ctrl_img: "/path/to/start_frame.jpg"