Best Local AI Video Model 2026: LTX-2.3 vs WAN 2.2 vs HunyuanVideo 1.5

Picking the best local AI video model between LTX-2.3, WAN 2.2, and HunyuanVideo 1.5 comes down to three questions: how much VRAM do you have, what does your video need to look like, and do you need audio. All three do both text-to-video (T2V) — generating a clip from a prompt alone — and image-to-video (I2V) — animating a reference photo — so this comparison covers both modes rather than treating them as separate decisions. It runs all three through the same categories — VRAM, speed, realism, stylized output, camera motion, physics, audio, and licensing — so you pick the right one before downloading 20GB+ of the wrong model.

All three run natively in ComfyUI with official or community-maintained nodes. If you haven't installed ComfyUI yet, install ComfyUI first before downloading any of the models below.

What You Need Before You Start

Check your GPU's VRAM before picking a model — this decides which one is even usable on your hardware, before quality enters the conversation. VRAM needs are the same whether you're generating text-to-video or image-to-video.

Quantization, in plain terms, is a compressed version of a model file — GGUF and FP8 formats trade a small amount of quality for a much smaller file size and lower VRAM use. All three models publish quantized versions.

For a full GPU buying guide across all three models — which card to buy at each budget, running on 8GB VRAM, fixing "CUDA out of memory" errors, and AMD/Apple Silicon/cloud rental options — see the VRAM & Hardware Guide for AI Video Generation.

LTX-2.3 vs WAN 2.2 vs HunyuanVideo 1.5 at a Glance

A fast top-level scan before the category-by-category breakdown below. Note that all three support both text-to-video and image-to-video — this isn't a text-to-video-only comparison.

Which Model Is Fastest?

LTX-2.3 is roughly 2–3x faster than WAN 2.2 or HunyuanVideo 1.5 at comparable settings, in both T2V and I2V. On an RTX 4090 with the FP8 distilled model at 8 steps, LTX-2.3 generates a 5-second clip in 20–40 seconds.

HunyuanVideo 1.5 sits in the middle — its smaller 8.3B parameter count (versus the older 13B HunyuanVideo) is specifically what made single-GPU generation practical, but it's still slower than LTX-2.3's distilled pipeline.

WAN 2.2 is the slowest of the three at standard settings — a full 14B render with 20+ steps can take 40–50 minutes for a single 5-second clip, for either mode. With the community Lightning LoRA — a distilled step-reduction LoRA that cuts sampling to 4 steps per expert model — that drops to roughly 1–3 minutes, closing most of the speed gap.

See Each Model's Own Text-to-Video Showcase

Before the matched-prompt breakdown below, here's each model's own T2V demo clip from its dedicated setup guide — these use different prompts per model, so treat them as "what this model looks like at its best," not a direct comparison.

Image-to-Video Compared: Animating a Reference Photo

This is the category most text-to-video-only comparisons skip — but all three models here ship a dedicated I2V pipeline, and for a lot of real workflows (bringing a product photo or character portrait to life) it matters more than T2V does. Each model below animates its own reference photo from its dedicated setup guide — not the same source image across all three — so read this as "what each model's I2V looks like at its best," the same honest framing as the T2V demos above.

LTX-2.3 — input photo

WAN 2.2 — input photo

HunyuanVideo 1.5 — input photo

On quality: WAN 2.2 generally holds up best for photorealistic I2V — skin texture and hair stay consistent frame to frame, particularly on close-up portraits. LTX-2.3 is close behind and is the only one of the three that can add synced audio to an animated image in the same generation pass. HunyuanVideo 1.5 does well with natural-feeling motion once the reference image and prompt are aligned, though its CLIP Vision step is an easy one to misconfigure — see the dedicated guide's troubleshooting section if your output ignores your input image.

Which Model Has the Best Video Quality?

"Best quality" isn't one number — it splits by what you're actually generating. Each section below uses the exact same prompt run through all three models, in text-to-video mode unless noted.

Realistic People and Photorealism

WAN 2.2 is generally considered the strongest for photorealistic human subjects — skin texture, hair, and lighting hold up best in close-up shots, in both T2V and I2V. HunyuanVideo 1.5 is close behind, with community testing praising its texture consistency, particularly in fine-tuned variants. LTX-2.3 is competitive at medium and wide shots, but its VAE architecture shows more softness in extreme close-ups than the other two.

A cinematic medium shot of a woman in her late 20s sitting at a cozy café beside a rain-streaked window, gently holding a warm cup of coffee with both hands. Raindrops slowly slide down the glass while soft afternoon light filters through the clouds. She slowly looks up toward the window, a subtle smile appears on her face, then she lowers her gaze back to the cup. Steam rises naturally from the coffee as she softly blinks. The camera performs a slow dolly-in with subtle handheld motion, creating an intimate, peaceful atmosphere. Natural movement, realistic facial expressions, gentle hair motion, cinematic lighting

Anime and Stylized Video

LTX-2.3 tends to produce the strongest results for stylized, animated, and non-photorealistic content — motion graphics, illustration styles, and artistic camera work. If your channel or client work leans stylized rather than photoreal, this is where LTX-2.3's speed advantage compounds: more iterations in the same amount of time.

An anime girl stands alone on a rooftop at sunset, her silver hair and jacket flowing gently in the evening wind. She takes a few slow steps toward the edge, pauses, and watches the sun disappear below the skyline. Birds fly across the distant sky while clouds slowly drift overhead. The camera starts behind her, slowly circles around to the front, and finishes with a close-up of her peaceful smile. Smooth cinematic camera movement, expressive animation, natural hair and clothing motion, warm golden evening light. 

Camera Motion and Cinematography

LTX-2.3 follows detailed camera instructions — dolly, pan, push-in, lens type — closely, and specifying exact cinematography terms in the prompt produces noticeably different results. WAN 2.2 also handles camera movement well and holds background elements steady during complex foreground motion. HunyuanVideo 1.5 is generally solid but can show slight warping on very long, complex camera moves.

The camera glides slowly through an ancient stone corridor toward a weathered wooden door. Warm candlelight dances across the rough stone walls while dust drifts gently through the air. As the camera reaches the doorway, it smoothly arcs to the right, revealing a wooden table holding a softly flickering candle beside an old parchment. The scene remains calm and continuous with slow cinematic movement, subtle environmental animation, warm ambient lighting, and a mysterious atmosphere.

Natural Motion and Physics

WAN 2.2 currently leads here. Fluid dynamics — water, smoke, fire — cloth simulation, and physical object interactions look more physically grounded than in LTX-2.3 or HunyuanVideo 1.5 at equivalent settings, in both T2V and I2V. If your scene depends on believable physical motion rather than a static or slow-moving subject, WAN 2.2 is worth the extra generation time (or the Lightning LoRA if you want that quality without the wait).

HunyuanVideo 1.5 was the physics leader in this comparison when it first released, but AI video moves fast — LTX-2.3 and WAN 2.2 have both had major updates since, and at roughly a year old, HunyuanVideo 1.5 has been passed on this front too. It's not a bad model, it's just no longer the newest one in the room.

Native Audio

LTX-2.3 is the only model of the three that generates audio synced to the video in the same pass — not added afterward as a separate step, for text-to-video or image-to-video. Neither WAN 2.2 nor HunyuanVideo 1.5 generates audio natively; if you need sound with either of those, you're adding it in a separate pipeline step outside ComfyUI's video generation.

Licensing — What You Can Actually Use Commercially

This matters more than most comparisons mention, and it applies the same whether you're generating T2V or I2V output. WAN 2.2 is released under the Apache 2.0 license — the most permissive of the three, with no meaningful commercial restrictions.

HunyuanVideo 1.5 uses the Tencent Hunyuan Community License — not Apache 2.0 or MIT. It allows personal, research, and most commercial use, but it explicitly does not apply in the EU, UK, or South Korea, and any product built on it that reaches over 100 million monthly active users needs a separate license from Tencent.

LTX-2.3's weights are published on HuggingFace under an open license permitting personal and commercial use. Terms can update, so check the current license file on the HuggingFace repo before publishing commercially — this applies to all three models, not just LTX-2.3.

Winner by Category

So Which One Should You Actually Use?

Nothing stops you from running all three — the models don't conflict, and many workflows keep two installed: one for speed (LTX-2.3) and one for a specific quality need (WAN 2.2 or HunyuanVideo 1.5), regardless of whether you're mostly doing T2V, I2V, or both.

Full Setup Guides for Each Model

Each of these has a complete, tested ComfyUI setup guide covering both text-to-video and image-to-video, with exact model downloads, folder structure, and settings.

Troubleshooting: Common Setup Issues Across All Three

"This node type does not exist" (red nodes on any of the three workflows)

You're missing a custom node the workflow depends on — this happens most often with HunyuanVideo 1.5's CLIP Vision node (needed for I2V specifically), LTX-2.3's GGUF workflow, or WAN 2.2's WanVideoWrapper nodes.

1. Click Manager in the top menu.
2. Click Install Missing Custom Nodes.
3. Restart ComfyUI completely, not just a browser refresh.

Black frames or no output on any model

Almost always a missing or misplaced text encoder file. Each of the three needs its text encoder(s) in models/text_encoders/ exactly as named in its dedicated guide above — double-check the filename, not just the folder.

Image-to-video output ignores your reference photo

This is specific to I2V and usually means a wiring problem, not a bad image. On HunyuanVideo 1.5, confirm CLIPVisionLoader and CLIPVisionEncode are actually connected into the HunyuanVideo15ImageToVideo node — if that link is missing, ComfyUI still generates a video, it just won't use your photo. On LTX-2.3 and WAN 2.2, confirm you loaded the I2V-specific model file, not the T2V one, into the checkpoint/model loader.

"CUDA out of memory" during generation

Your VRAM doesn't match the model variant you downloaded.

1. Confirm you downloaded the quantized (FP8 or GGUF) version, not full BF16/FP16, if you're under 16GB.
2. Lower resolution or frame count first before switching models entirely.
3. Close other GPU-heavy applications and retry.

For anything not covered here, see the full ComfyUI troubleshooting guide.

Frequently Asked Questions

Published: 2026-07-17 · Last updated: 2026-07-18 · Models compared: LTX-2.3 (Lightricks), WAN 2.2 (Alibaba), HunyuanVideo 1.5 (Tencent) · Both text-to-video and image-to-video covered for each model