R_REDDYX.XYZ
Voice and Media2026-06-12

Video generation with neural networks in 2026: Sora, Runway, Kling and open source

Video generation with neural networks in 2026: Sora, Runway, Kling and open source
TL;DR: Generation of video by neural network in 2026 split into three camps: closed flagship models (Sora, Runway, Kling) provide cinematic picture and long clips, while open source (Wan, HunyuanVideo, LTX-Video, Mochi) caught up in quality of short clips and runs locally for free. For production choose closed APIs, for pipelines and experiments — open weights on your own GPU.

WHAT HAS CHANGED OVERALL BY 2026

Two years ago, AI video meant five-second glitchy clips with floating fingers and face morphing. In 2026 the picture is different. Flagship models confidently hold scene physics for 10-20 seconds, understand camera ('dolly-in', 'orbit around object'), synchronize lips with speech and generate sound together with the image. Video generation has stopped being a demo toy and has entered real pipelines: advertising, pre-visualization, motion graphics, b-roll for YouTube.

The main shift is not so much in the quality of individual flagships, but in the fact that open source has stopped being laughable. Models with open weights now produce results that a year ago were considered exclusive to closed APIs. This breaks the old logic of 'pay or nothing'.

SORA: OPENAI FLAGSHIP

Sora from OpenAI remains the benchmark for 'understanding what is happening in the scene'. The model is strong where long narrative consistency matters: the object does not teleport between frames, shadows fall correctly, water behaves like water. The second generation of Sora added synchronized sound generation and noticeably improved control over characters — you can lock the hero's appearance and transfer it between scenes.

Weak points honestly exist. The model is expensive in credits, access goes via subscription and web interface, and the API for programmatic access is surrounded by limits and moderation. For quick iterations like 'generated 50 variants in an evening' this is a poor fit — you'll hit the quota and your wallet.

  • Strength: narrative cohesion, physics, long shots, built-in sound.
  • Weakness: price, strict moderation, limited programmatic access.
  • For whom: those who need one perfect shot, not a hundred drafts.

RUNWAY: TOOL FOR THE EDITOR

Runway with the Gen-4 lineup bet not on 'wow generation from text', but on controllability. Here is the industry's best set of tools around video: motion brush (you specify which part of the frame moves), camera trajectory control, image-to-video preserving style, video inpainting, frame extension. Runway is less 'magic from the prompt', more 'workstation'.

For those who actually edit, Runway is often more convenient than Sora: predictability matters more than peak quality. Reference-based generation (you give a character image and get it in motion) works stably, which is critical for branded content where the hero must look the same across ten clips.

KLING: CHINESE HEAVYWEIGHT

Kling from Kuaishou became the main surprise. According to community feelings, it often delivers the most realistic human movement and best lip-sync among closed models, while being cheaper than Sora. Kling is strong in human anatomy and dance — where other models break on fast limb movement. There are separate modes for faces, lip-sync to audio, and extending the clip beyond a minute.

Downside — access and privacy. The model is cloud-based, servers in China, for sensitive content this is a showstopper. However, by price-to-realism ratio in 2026, Kling is one of the best options for people-centric content.

OPEN SOURCE: WAN, HUNYUANVIDEO, LTX, MOCHI

The most interesting developments are happening in open weights. Here are several families, each suited to a different task:

  1. Wan (Alibaba) — currently considered the strongest open text-to-video and image-to-video model in quality. There are versions of different sizes for different hardware.
  2. HunyuanVideo (Tencent) — a large model with very 'cinematic' aesthetics, fine-tunes well to your style via LoRA.
  3. LTX-Video (Lightricks) — speed king. It generates a short clip faster than real-time on a top consumer card. For iterations and pipelines — gold.
  4. Mochi (Genmo) — strong foundation for motion, fully open Apache 2.0 license.

The key advantage of open models: they are free to use, run locally (privacy), support fine-tuning and can be embedded in ComfyUI nodes. Fresh releases and forks of these models are conveniently tracked in the REDDYX catalog — weights and repositories are updated literally every week.

COMPARISON TABLE

ModelTypeStrengthClip lengthLocalPrice
SoraClosedScene coherence, soundup to ~20 secNoSubscription, expensive
Runway Gen-4ClosedControl, tools~10 sec + extendNoCredits, medium
KlingClosedHuman realism, lip-syncup to ~1 min+NoCredits, cheap
WanOpen sourceGeneral quality~5 sec/segmentYesFree
HunyuanVideoOpen sourceCinematic style, LoRA~5-6 secYesFree
LTX-VideoOpen sourceSpeed~5-10 secYesFree

The length and quality figures are generalized, based on community measurements and authors' statements; specific values depend on the version, resolution, and sampler settings.

LAUNCHING OPEN SOURCE LOCALLY

Practice is more important than theory. Below is a minimal working example of video generation on an open model using the diffusers library. A GPU with sufficient VRAM is required (realistically — 16 GB and higher, with optimizations you can go lower).

import torch
from diffusers import DiffusionPipeline
from diffusers.utils import export_to_video

# грузим открытую видеомодель (пример на HunyuanVideo)
pipe = DiffusionPipeline.from_pretrained(
    "hunyuanvideo-community/HunyuanVideo",
    torch_dtype=torch.bfloat16,
)

# экономим VRAM: выгружаем слои на CPU по мере надобности
pipe.enable_model_cpu_offload()
pipe.vae.enable_tiling()

prompt = (
    "cinematic shot of a red fox running through snowy forest, "
    "golden hour light, shallow depth of field, 35mm film look"
)

output = pipe(
    prompt=prompt,
    height=544,
    width=960,
    num_frames=61,      # ~2.5 сек при 24 fps
    num_inference_steps=30,
    generator=torch.Generator("cuda").manual_seed(42),
).frames[0]

export_to_video(output, "fox.mp4", fps=24)
print("Готово: fox.mp4")

Want to stay ahead with these tools — subscribe to Telegram channel REDDYX AI: new repositories and reviews every 30-60 minutes.

New repositories every 30 minutes

REDDYX AI scans GitHub 24/7 and ships the best AI/ML/Web3 projects to Telegram.

Join on Telegram

← All articles

Video generation with neural networks in 2026: Sora, Runway, Kling and open source | REDDYX AI