Skip to content
FlixMLFlixML Docs

Wan 2.2 Image-to-Video

Author-sourced reference for the Wan 2.2 I2V model: how the architecture works, how motion is driven, and how the model authors say to prompt it. This is not a description of our Studio wiring — it distills the official Wan-Video / Alibaba / ComfyUI documentation so we prompt and configure the model the way its creators intended.

Sources are cited inline. Primary references:

1. What the model is

Section titled “1. What the model is”

Wan 2.2 is Alibaba’s open-weights (Apache 2.0) video generation family. The image-to-video variant is Wan2.2-I2V-A14B: it takes a starting image (+ optional text prompt) and animates it into video while preserving the image’s content, subject, scene, and style.

Three model variants ship:

Variant Params Role
Wan2.2-TI2V-5B 5B Hybrid t2v + i2v, single model. Fits ~8GB VRAM with offloading. Fast prototyping.
Wan2.2-I2V-A14B 14B (MoE) Dedicated image-to-video. Highest i2v quality.
Wan2.2-T2V-A14B 14B (MoE) Dedicated text-to-video.

Source: ComfyUI tutorial model table — docs.comfy.org/tutorials/video/wan/wan2_2

The 14B models are the ones with the two-expert architecture below. The 5B hybrid is a single dense model (no high/low-noise split).

2. The MoE architecture — why there are two model files

Section titled “2. The MoE architecture — why there are two model files”

The 14B models use a Mixture-of-Experts (MoE) design on a flow-matching Diffusion Transformer (DiT) backbone. This is the single most important thing to understand about Wan 2.2, because it’s why i2v ships as two .safetensors files (high_noise + low_noise).

  • Two specialized experts, ~14B params each (~27B total), but only one is active per denoising step — so per-step compute/memory stays close to a dense 14B model.
  • The experts are temporally specialized by noise level, not by content. They split the denoising trajectory in two.

Source: wan.video/blog/wan2.2; Wan-AI/Wan2.2-I2V-A14B model card; deeplearning.ai The Batch.

High-noise expert (early denoising)

Section titled “High-noise expert (early denoising)”
  • Active in the early steps (high noise, low signal-to-noise ratio).
  • Owns global structure: composition, subject placement, coarse motion, scene layout — the high-level “plan” of the shot as it emerges from pure noise.

Low-noise expert (late denoising)

Section titled “Low-noise expert (late denoising)”
  • Active in the later steps (low noise, high SNR).
  • Owns refinement: textures, fine detail, edges, lighting, and temporal consistency.

How the switch happens (SNR routing)

Section titled “How the switch happens (SNR routing)”
  • Routing is by signal-to-noise ratio (SNR), which falls monotonically as denoising progresses.
  • A boundary timestep t_moe is defined where SNR = SNR_min / 2.
    • t ≥ t_moe (high noise) → high-noise expert.
    • t < t_moe (low noise) → low-noise expert.

Source: Wan-AI/Wan2.2-I2V-A14B model card; wan.video/blog/wan2.2.

Practical implication for us: in a ComfyUI graph you load both diffusion models and sample in two stages — high-noise model for the first block of steps, low-noise model for the remainder. Getting the step split / boundary right matters: motion and structure are decided in the high-noise stage, so that stage is what governs how much and what kind of motion you get.

3. How motion is driven

Section titled “3. How motion is driven”

Two inputs shape motion in i2v:

  1. The starting image anchors identity, framing, scene, and style. The model animates from it — it does not re-invent the subject.
  2. The text prompt drives what moves and how the camera moves. For i2v the authors are explicit that the prompt’s job is motion + camera, not describing the subject again (the image already did that).

The high-noise expert is where coarse motion and layout are established during early denoising; the low-noise expert then keeps that motion temporally consistent while sharpening detail. So motion “amplitude” is largely a high-noise-stage phenomenon.

Wan 2.2 was specifically trained for large-scale complex motion and cinematic camera control — the authors call out improved motion smoothness/controllability and professional camera-language adherence as headline features.

Source: ComfyUI tutorial “Model Highlights”; wan.video/blog/wan2.2.

4. Official prompting formulas (from the model authors)

Section titled “4. Official prompting formulas (from the model authors)”

These come straight from Alibaba Cloud Model Studio’s prompt guide — the authoritative source for how to prompt this model family.

Source: help.aliyun.com/en/model-studio/text-to-video-prompt

Image-to-video formula (use this one for i2v)

Section titled “Image-to-video formula (use this one for i2v)”
Prompt = Motion + Camera movement

Because the image already defines entity, scene, and style, the i2v prompt should focus only on:

  • Motion — what the elements do. Describe amplitude, speed, effect. Use modifiers like “slowly,” “quickly,” “gently,” “violently.” e.g. “the subject slowly turns her head and smiles,” “fabric sways in a gentle breeze.”
  • Camera movement — explicit camera instructions: “camera pushes in,” “camera moves left,” “slow pan right,” “orbit around the subject.” To hold the camera still, say “fixed camera” — if you omit camera direction the model tends to invent a default move.

For reference — the text-to-video formulas (context, not for i2v)

Section titled “For reference — the text-to-video formulas (context, not for i2v)”
  • Basic: Entity + Scene + Motion
  • Advanced: Entity + Scene + Motion + Aesthetic control + Stylization
    • “Aesthetic control” = light source, lighting, shot size, camera angle, lens, camera move.
    • Most weak prompts cover entity + motion but skip aesthetic control → you get a static camera in an undefined void. (Authors’ own words.)

Camera language (from the guide)

Section titled “Camera language (from the guide)”

The guide maps camera moves to their on-screen effect — use professional cinematography vocabulary for predictable results:

  • Push-in → emphasis on the subject
  • Pull-out → reveals scale and surroundings
  • Tracking shot → viewer moves alongside the subject
  • Orbit → circles the subject
  • Fixed camera → stillness and focus

Other supported moves: pan (left/right), tilt (up/down), dolly, crane, handheld, zoom.

Prompt hygiene

Section titled “Prompt hygiene”
  • Be explicit about camera — omission yields default motion.
  • Keep prompts focused; combine motion modifiers with clear camera intent.
  • prompt_extend (Alibaba API / a Qwen model) can auto-expand a short prompt into a full structured one if you want the model to fill in cinematic detail.

5. ComfyUI wiring the authors ship (14B i2v)

Section titled “5. ComfyUI wiring the authors ship (14B i2v)”

The official ComfyUI i2v workflow loads both experts and samples in two stages:

  • Load Diffusion Model #1 → wan2.2_i2v_high_noise_14B_*.safetensors
  • Load Diffusion Model #2 → wan2.2_i2v_low_noise_14B_*.safetensors
  • Text encoder: umt5_xxl (UMT5) — Wan uses UMT5 for text conditioning.
  • VAE: wan2.2_vae.safetensors (I2V image conditioning uses the 3D VAE; compatible with the Wan2.1 VAE).
  • The input image feeds the i2v latent node; length sets total frames.

Source: docs.comfy.org/tutorials/video/wan/wan2_2 (14B I2V section); model card (UMT5 / 3D VAE).

For low-VRAM boxes, GGUF quantized builds exist (e.g. bullerwins/Wan2.2-I2V-A14B-GGUF) and speed LoRAs (Lightx2v 4-step) are published for faster sampling — cited in the same ComfyUI tutorial. These are third-party accelerations, not core model behavior.

6. TL;DR for prompting our i2v shots

Section titled “6. TL;DR for prompting our i2v shots”
  1. Don’t re-describe the subject — the image already carries it.
  2. Write motion first, camera second: what moves + how the camera moves.
  3. “Motion” means REAL motion — walking, running, dancing, jumping, workouts, exercises, sports: big physical action. Describe the full movement with strong verbs. Do not default to subtle / slightly / gently / slowly / drifting / breeze — that vocabulary yields a near-still “breathing photo,” which is the wrong result unless motion is explicitly not wanted.
  4. Ambient/subtle language is only for talking-head or deliberately-still shots (lip-sync clips where the character just speaks). Know which mode you’re in before writing the prompt.
  5. Always state a camera instruction — even “fixed camera” — or the model picks one for you.
  6. Two-stage denoise: coarse motion/structure is set in the high-noise stage, so its step count (steps_high) is the lever for how much motion you get. Real motion needs the full step count (default 15/15 = 30); a 4-step Lightning preview kills the motion.