Prompt describes a scene, not an action; or too few steps on the high-noise expert. Add motion verbs and camera moves; give the high-noise expert half the steps.

Trends

.ai

i migliori modelli di IA open source

Wan 2.2 video: text-to-video and image-to-video setup, settings and LoRAs

Q: Out of memory at the first sampler

fp16 14B weights on a 24 GB or smaller card. fp8 or GGUF Q4 pair, fewer frames, 480p, --lowvram; or the 5B model.

Q: Decode fails after sampling succeeded

VAE decode is the memory peak for video. Wan VAE decode in tiles (the tiled option on the decode node), or fewer frames.

Argomento: Modelli video

Di Captain

Pubblicato il 2026-05-24

Getting first results from Wan 2.2: which model size to pick (5B or 14B), the high-noise / low-noise expert pair, resolution and frame counts that work, the lightning LoRAs that cut renders to four steps, prompt structure for motion, and how to run it hosted when the card is too small.

Questa pagina non è ancora tradotta, quindi è mostrata in inglese.

Panoramica

Wan 2.2

is Alibaba's open video family and the most downloaded video model on this site. It comes as a 5B hybrid text-and-image-to-video model built for consumer cards (720p on 8 to 12 GB), and as 14B text-to-video and image-to-video models that use two experts: a high-noise model that lays out motion and composition in the first steps and a low-noise model that adds detail in the last steps. The 14B pair produces the cinematic results people share; the 5B model is the one to learn on.

The community has trained hundreds of

Wan

LoRAs

: camera moves (orbit, dolly, crash zoom), motion styles, characters, effects, and the lightning / distillation LoRAs that reduce 30-step renders to 4 to 8 steps with little loss. The Wan family page ranks them by heat; the lightning LoRAs are nearly always at the top.

Hosted options are a good fit for video because renders are long and VRAM-hungry:

PirateDiffusion

runs Wan workflows from Telegram with /wf /run, and

BitVector

keeps the Wan graphs preinstalled with the LoRAs in place. Both are linked in the Run box of every Wan model page.

Riferimento

Nome	Tipo	Cos'è
Wan2.2-TI2V-5B	model	One model for text- and image-to-video, 720p at 24 fps, 121 frames max. fp16 10 GB, fp8 5 GB. 8 to 12 GB cards.
Wan2.2-T2V-A14B / I2V-A14B	model pair	High-noise and low-noise experts, 14B each (27B total, 14B active). 480p and 720p. fp8 about 14 GB each; GGUF Q4 about 8 GB each.
umt5-xxl text encoder	file	fp8 scaled (6 GB) or fp16 (11 GB). Shared by all Wan models.
Wan 2.1 / 2.2 VAE	file	The 5B model uses the 2.2 VAE; the 14B models use the 2.1 VAE. Mixing them produces garbage.
Resolution	setting	480x832 or 832x480 for fast tests; 720x1280 / 1280x720 for finals. Multiples of 16.
Frames	setting	4n+1: 33, 49, 65, 81. 81 frames at 16 fps is five seconds.
Steps / shift / CFG	setting	20 to 30 steps, shift 5 to 8, CFG 3.5 to 5 without lightning. With lightning LoRAs: 4 to 8 steps, CFG 1, shift 5.
Lightning / distill LoRA	LoRA	Separate LoRAs for the high-noise and low-noise experts; load each on its own model. Weight 1.0.
Prompt	text	Subject, action, camera, lighting, style, in that order. Describe motion verbs explicitly. Negative prompt is used by Wan (unlike Flux ).

Nessuna installazione: esegui i modelli IA nel cloud

BitVector Prism è il modo più semplice per iniziare: scegli un modello, scrivi un prompt e genera in una web app pulita, senza configurare nulla. BitVector è disponibile anche su Discord e sul web (SpyGlass).

Prova BitVector

Passo dopo passo

Install the loaders:
ComfyUI
core supports Wan 2.2; add ComfyUI-GGUF for quantised files. Update ComfyUI first.
Download for your card: 5B fp8 plus the 2.2 VAE and umt5 fp8 on 8 to 12 GB; the 14B fp8 pair plus the 2.1 VAE on 24 GB; the 14B GGUF Q4 pair on 12 to 16 GB.
Load the official template (Workflow > Browse Templates > Video > Wan 2.2) and replace the loaders with your files. For GGUF swap Load Diffusion Model for Unet Loader (GGUF).
For 14B: two KSampler Advanced nodes. The first runs the high-noise model for steps 0 to N/2 (return with leftover noise), the second runs the low-noise model from N/2 to N. The template wires this already.
First render: 480x832, 33 frames, 20 steps, CFG 4, shift 8, prompt "a corgi runs along a beach at sunset, camera tracks alongside, golden light, shallow depth of field". Expect 2 to 6 minutes.
Add the lightning LoRAs from the Wan family page (one for each expert), set steps to 6 (3 + 3), CFG 1. Renders drop to under a minute at 480p.
Image-to-video: load the I2V pair, feed the picture into WanImageToVideo, keep the prompt about motion and camera rather than describing the picture.
Finals: 720p, 81 frames, interpolate to 32 fps with RIFE or FILM, upscale with a video
upscaler
from the
Upscalers
family if needed.
Too slow or too big: run the same workflow on BitVector, or on PirateDiffusion with /workflow /show:wan to read the fields and /wf /run:wan your prompt.

Esempi

Text-to-video prompt

A fisherman in a yellow raincoat hauls a net onto a small wooden boat in heavy rain, waves rock the boat, the camera is handheld and close, grey stormy light, cinematic, 35mm film grain

Negative: overexposed, static, blurry details, subtitles, worst quality, low quality, deformed, extra fingers, still image

Image-to-video prompt (motion only)

The woman turns her head toward the camera and smiles, wind moves her hair, slow push-in, soft daylight

PirateDiffusion

/workflow /show:wan-i2v

/wf /run:wan-i2v /initimage:Iabc123 the woman turns her head toward the camera and smiles, wind moves her hair, slow push-in

Consigli

Describe motion with verbs and the camera with film words (slow dolly in, handheld, static wide shot). Wan follows camera language unusually well.
Keep the negative prompt: the official one lists overexposure, static, blurry details, subtitles, worst quality, extra fingers, and it helps.
For I2V, generate the first frame with Flux,
Qwen
or
Z-Image
at the target aspect ratio; a sharp, well-lit still produces a better clip than any prompt.
Frame count costs more than resolution: 81 frames at 480p is cheaper than 49 frames at 720p.
Motion LoRAs for Wan are tagged by the expert they target; a low-noise-only LoRA adds detail, a high-noise LoRA changes motion.
Seed matters a lot in video; keep the seed fixed while you tune the prompt, then vary it for takes.
Save the workflow with your LoRA set as a template; Wan graphs are long and easy to break.

Match version and size; the model page names both.

Nessuna installazione: esegui i modelli IA nel cloud

PirateDiffusion è solo su Telegram ed è pensato per i professionisti: migliaia di modelli, LoRA e workflow tramite comandi in chat, con generazione illimitata a prezzo fisso.