.ai

les meilleurs modèles d'IA open source

Rechercher

AI video generation for beginners: image-to-video, text-to-video and the first clip with Wan, LTX-2 and H3

Q: Camera jumps or the scene cuts

Two camera instructions, or "cut to" words in the prompt. One camera move; put "jump cut" in the negative.

Q: Out of memory at 720p

Frames times resolution exceed VRAM. Drop to 480p or fewer frames, offload the text encoder; see the VRAM for AI video guide.

Q: Motion LoRA does nothing

Loaded on the wrong expert or the wrong model size. Check the model page for 5B/14B and high/low noise; load accordingly with its trigger word.

Sujet : Modèles vidéo

Par Lookout

Publié le 2026-04-29

A first week with open video models: the difference between text-to-video and image-to-video and why to start with the latter, the three model families worth learning (Wan 2.2, LTX-2, MiniMax H3), frame counts and resolutions that work, how to write a motion prompt, what to expect from a 16 or 24 GB card against the cloud, and how to turn five-second clips into something longer.

Cette page n'est pas encore traduite ; elle est affichée en anglais.

Aperçu

Video models are image models with a time axis: the same denoising loop runs on a block of frames at once, so everything you know about prompts, seeds and denoise carries over, and so do the costs, multiplied by the frame count. The practical consequences for a beginner are three. Start with image-to-video (I2V), because a good still from an image model fixes composition and style and leaves the video model only the motion to invent. Keep clips short (about five seconds), because that is what the models were trained on. And expect to run the large models on the cloud unless you own a 24 GB card.

Three open families cover most needs in 2026.

Wan 2.2

(Alibaba) is the all-rounder: a 5B model that runs on 12-16 GB and a 14B two-expert model with the best open quality, plus a huge

LoRA

ecosystem for motion and style.

LTX-2

(Lightricks) is fast and generates synced audio but wants 24-32 GB.

MiniMax H3

is the audio-video model with dialogue and sound effects, mostly used through hosted workflows; the H3 prompting guides on this site by Mark White go deep on it.

Hunyuan

Video 1.5 is a lighter alternative to

Wan

14B.

The catalog lists each video model and its LoRAs with the base and size they target (5B or 14B, high-noise or low-noise expert), and the model pages show example clips with their prompts. The Run button opens the model on

PirateDiffusion

, where a workflow command produces the clip from a reply to your image, or on

BitVector

Référence

Nom	Type	Rôle
I2V (image to video)	mode	Animate a still. Best first mode: composition is already right; the prompt describes motion only.
T2V (text to video)	mode	From a prompt alone. More variety, less control; use for ideas and B-roll.
FLF2V / first-last frame	mode	Two keyframes; the model fills the motion between. Controlled transitions.
Frames and fps	setting	Wan: 81 frames at 16 fps (5 s); LTX-2: 97-121 frames at 24-25 fps; H3: 145 frames at ~24 fps. Memory scales with frames.
Resolution	setting	Wan 2.2: 832x480 or 480x832 (fast), 1280x720 (24 GB+); LTX-2 up to 1080p+ with 32 GB. Portrait for people, landscape for scenes.
Motion prompt	prompt	One subject motion plus one camera motion: "she turns toward the window and smiles; slow push-in". Avoid listing scene details already in the image.
Steps / CFG / shift	setting	Wan 2.2 14B: 20 steps split 10/10 across experts, CFG 3.5, shift 5; 5B: 20-30 steps, CFG 5; LTX-2: 25-30 steps with its own schedule; distilled LoRAs (lightx2v) bring Wan to 4-8 steps at CFG 1.
Motion LoRAs	file	Camera moves, dances, gestures, physics; matched to 5B/14B and to the expert. Pages here name the target.
Interpolation and upscale	post	RIFE or FILM to 32-60 fps; frame-wise ESRGAN or SeedVR for resolution.

Your first clip without a GPU

BitVector Prism animates a still with the Wan models from this guide in the browser: upload the image, write the motion sentence, download the clip. Nothing to install, works from a laptop.

Essayer BitVector

Pas à pas

Make or pick a still: a 1024x1024 or 832x1216 render with a clear subject and simple background. Faces larger than a tenth of the frame animate best.
Open a Wan 2.2 I2V model page on this site, copy the example motion prompt pattern, and press the Run button (PirateDiffusion: reply to your image with the /workflow command; BitVector: upload and pick the model).
Write a motion prompt of one sentence for the subject and one for the camera. Nothing about clothes, colours or lighting; the image has those.
Generate at 832x480 (landscape) or 480x832 (portrait), 81 frames, default steps. Watch the whole clip; judge motion, not sharpness.
Iterate on the motion prompt only: slower verbs ("slowly", "gently") reduce flailing; a single camera instruction prevents jumps.
Locally, with a 16 GB card: Wan 2.2 5B fp8 or GGUF, 832x480, 49-81 frames, text encoder on CPU. With 24 GB: 14B fp8 at 480p, 720p with the encoder offloaded. See the
VRAM for AI video
guide.
Longer pieces: take the last frame of clip one as the start image of clip two with a continuing motion prompt; join in an editor; interpolate to 30 fps and upscale at the end.
Add sound: LTX-2 and H3 generate audio with the clip; for Wan, add music or effects in the editor.

Exemples

Motion prompt for I2V

Positive: she lifts the coffee cup, takes a sip and looks out of the window; slow push-in, shallow depth of field, natural motion

Negative (Wan): static, blurry, distorted face, extra fingers, jump cut, text

PirateDiffusion workflow command

/workflow /run:wan22-i2v she lifts the coffee cup and looks out of the window, slow push-in /length:81 (as a reply to the still)

ComfyUI Wan 2.2 5B on 16 GB

Astuces

Motion verbs matter more than adjectives: "turns, walks, lifts, pours, waves" each produce distinct, reliable motion; "beautiful cinematic" produces nothing.
One camera move per clip: push-in, pull-back, pan left, orbit. Two moves give a wobble.
Portrait stills animate faces better; landscape stills animate environments better.
Wan LoRAs for the high-noise expert shape the motion; low-noise ones shape detail. Many packs ship both; load each on its expert.
Distilled LoRAs (lightx2v, FastWan) cut Wan from 20 steps to 4-6 at CFG 1 with little quality loss; the model pages list them.
Expect 3-10 minutes per clip on a 4090 for Wan 14B and about 90 seconds for LTX-2; the cloud returns most clips in a minute or two.

Dépannage

The clip is a still with a slight shimmer

Pourquoi cela arrive

No motion verb in the prompt, or a static LoRA weight too high.

Comment le corriger

Add a subject motion and a camera motion; lower LoRAs to 0.7.

Face melts or changes identity

Pourquoi cela arrive

Face too small, too many frames, or 480p on a detailed face.

Comment le corriger

Crop closer, 49-61 frames, or 720p on the cloud; a face-detailer pass per frame is a last resort.

Camera jumps or the scene cuts

Pourquoi cela arrive

Two camera instructions, or "cut to" words in the prompt.

Comment le corriger

One camera move; put "jump cut" in the negative.

Out of memory at 720p

Pourquoi cela arrive

Frames times resolution exceed

VRAM

Comment le corriger

Drop to 480p or fewer frames, offload the text encoder; see the VRAM for AI video guide.

Motion LoRA does nothing

Pourquoi cela arrive

Loaded on the wrong expert or the wrong model size.

Comment le corriger

Check the model page for 5B/14B and high/low noise; load accordingly with its trigger word.

Wan 2.2 14B, LTX-2 and H3 from Telegram

PirateDiffusion runs the large video models on server GPUs: reply to any image with a workflow command and the clip comes back in the chat. Unlimited generation on a fixed price, no 24 GB card needed.