Wan 2.2 TI2V 5B

The dense 5B member of Wan 2.2. One transformer does both text-to-video and image-to-video, working in the 16× spatial / 4× temporal latent space of the new Wan 2.2 VAE. Small enough to train on a single consumer GPU.

weights: Wan-AI/Wan2.2-TI2V-5B-Diffusers ↗
org: Wan-AI (Alibaba)
modality: video
tasks: text-to-video · image-to-video · text-to-image
license: Apache 2.0
released: 2025-07-28not verified
native output: 1280×704, up to 121 frames, 24 fpsnot verified
total params: 11.39B
model.arch: wan22_5b

Components

role	model	params	size	dtype	trained
Transformer	Wan 2.2 TI2V 5B DiT Wan-AI/Wan2.2-TI2V-5B-Diffusers/transformer diffusers.WanTransformer3DModel Stored as fp32 in the Diffusers repo, which ai-toolkit reads only for the config. The weights it loads are Comfy-Org/Wan_2.2_ComfyUI_Repackaged split_files/diffusion_models/wan2.2_ti2v_5B_fp16.safetensors (9,999,658,848 bytes).	5.00B 4,999,787,712	20.00 GB	fp32	yes
Text encoder	UMT5-XXL (encoder only) Wan-AI/Wan2.2-TI2V-5B-Diffusers/text_encoder transformers.UMT5EncoderModel Includes the 256k-token embedding table. ai-toolkit does not load this copy: it uses the config and tokenizer from ai-toolkit/umt5_xxl_encoder and the weights from Comfy-Org/Wan_2.1_ComfyUI_repackaged (umt5_xxl_fp8_e4m3fn_scaled or umt5_xxl_fp16).	5.68B 5,680,910,336	11.36 GB	bf16	no
Tokenizer	UMT5 SentencePiece (256k vocab) Wan-AI/Wan2.2-TI2V-5B-Diffusers/tokenizer transformers.T5TokenizerFast	—	—	—	—
VAE	Wan 2.2 VAE Wan-AI/Wan2.2-TI2V-5B-Diffusers/vae diffusers.AutoencoderKLWan Not the Wan 2.1 VAE. The 14B Wan 2.2 models still use the 2.1 VAE; only this one uses the 2.2 VAE.	704.7M 704,688,668	2.82 GB	fp32	no
total		11.39B	34.18 GB

Latent space

spatial

16×

temporal

4×

channels

patch

1×2×2

autoencoder: Wan 2.2 VAE
pixels per token: 32×32 × 4 frames
frame count: 4n + 1
notes: The VAE does a 2×2 pixel unshuffle before encoding (12 input channels = 3 × 2 × 2), then 8× downsampling, for 16× total. The first frame gets its own latent frame, which is why frame counts are 4n + 1.

input	latent (c×t×h×w)	tokens
768×768 × 121f	48×31×48×48	17,856	ai-toolkit sample default
1280×704 × 121f	48×31×44×80	27,280	native 720p
1024×1024	48×1×64×64	1,024	still image

Architecture

Blocks: 30
Hidden size: 3072 (24 heads × 128)
FFN size: 14336
Text conditioning: Cross-attention on UMT5 hidden states (4096-d), 512 tokens max
Image conditioning: None in the weights. I2V puts the clean first-frame latent in latent frame 0
Timesteps: Per token (expand_timesteps), so conditioned tokens can sit at t = 0
Objective: Rectified flow, shift 5.0
Norm / position: QK RMSNorm across heads, 3D RoPE

In AI Toolkit

model.arch: wan22_5b
UI label: Wan 2.2 TI2V (5B) (video)
model.name_or_path: Wan-AI/Wan2.2-TI2V-5B-Diffusers
source: extensions_built_in/diffusion_models/wan22/wan22_5b_model.py
toolkit/models/wan21/wan21.py
toolkit/models/wan21/wan_utils.py
toolkit/models/v2/_mixin.py
toolkit/models/v2/diffusion_models/wan.py
toolkit/models/v2/text_encoders/umt5.py
extensions_built_in/diffusion_models/ui.tsx
extra UI sections: sample.ctrl_img, datasets.num_frames, datasets.do_i2v, datasets.auto_frame_count, model.low_vram

UI defaults

quantize / quantize_te: true / true (qfloat8)
low_vram: true
noise_scheduler / sampler: flowmatch / flowmatch
timestep_type: weighted
sample size: 768×768, 121 frames, 24 fps
datasets.do_i2v: true
datasets.fps: 24
network.conv: disabled (linear LoRA only)

Specifics

Resolution: Buckets snap to multiples of 32 (16× VAE × 2×2 patch). Wan 2.1 uses 16.
Frame count: Sampling rounds num_frames down to 4n + 1.
Image-to-video: With do_i2v on, the first frame is encoded and placed in latent frame 0 at t = 0 and masked out of the loss (the mask is renormalized). Uses cached first-frame latents when available. Still images train as plain text-to-image.
Weight source: For a Hub repo id, ai-toolkit loads ComfyUI repackaged weights and uses the named repo only for configs: the transformer from wan2.2_ti2v_5B_fp16.safetensors, and UMT5 from umt5_xxl_fp8_e4m3fn_scaled or umt5_xxl_fp16 (ranked by the requested qtype; a local copy of either wins). The UMT5 config and tokenizer come from ai-toolkit/umt5_xxl_encoder unless name_or_path is a local folder with text_encoder/. model_kwargs.use_comfy_weights: false opts out. The text encoder is never trained.
VAE source: vae/ inside name_or_path (extras_name_or_path defaults to it).
Quantization: condition_embedder* and proj_out* stay in full precision.
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 (shift 5.0). UniPC is disabled until a diffusers regression is fixed. low_vram also turns on VAE tiling.
Not supported: split_model_over_gpus, assistant_lora_path, inference_lora_path, lora_path
Metadata base version: wan_2.2_5b

Example config

not verified

job: extensionconfig:  name: "my_wan22_5b_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: 121          fps: 24          do_i2v: true          resolution: [512, 768]      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/Wan2.2-TI2V-5B-Diffusers"        arch: "wan22_5b"        quantize: true        quantize_te: true        low_vram: true      sample:        sampler: "flowmatch"        sample_every: 250        width: 768        height: 768        num_frames: 121        fps: 24        guidance_scale: 4        sample_steps: 30        prompts:          - "a bear building a log cabin in the snow covered mountains"