TL;DR: Lightricks released LTX-2.5 on August 11 as open weights — a 22B audio-video model with day-one ComfyUI templates that generates synced sound and picture in one pass. The int8 build genuinely runs on a 24GB card: a community recipe measured 22.67 GiB resident on an RTX 4090 for 5-second clips. Below 24GB you’re leaning on weight streaming and patience; at 32GB an RTX 5090 turns out a 4-second 720p clip in about 25 seconds.

	RTX 5090 32GB	Used RTX 3090 24GB	16GB and below
Best for	Fastest local path: ~25s per 4s clip	Cheapest real seat at the table	Testing the waters, not daily use
Price (Aug 2026)	$4,381+ street, $4,700 Newegg median	$1,201–$1,299 fair range	you already own it
The catch	8s clips still spill into weight streaming (~3 min)	~2.5 min per 5s clip on the distilled model	int8 transformer alone is ~21.5GB — everything runs offloaded

Honest take: If you already own a 24GB card, LTX-2.5 is the best local video model you can actually run this month — download it tonight. Nobody should buy a GPU for it below the used-3090 tier; rent an hour of cloud 5090 time first and see whether the output fits your use case.

Local video generation has spent two years being the thing 24GB owners read about while 80GB owners did it. Wan’s full-quality tiers want datacenter VRAM, HunyuanVideo’s floor is similar, and the models that did fit consumer cards produced silent clips that needed a separate audio pass. LTX-2.5 is the first release where the headline model — not a cut-down variant — lands on hardware this site’s readers own, with the audio already in the file. This is the guide to what it actually needs.

What shipped on August 11, exactly

LTX-2.5 is a 22-billion-parameter diffusion transformer from Lightricks that generates video and synchronized audio in a single forward pass, released as open weights on Hugging Face with day-one native support in ComfyUI. Two open-weight versions exist: Dev, the full model, and Distilled, a faster checkpoint that samples in 8 steps at CFG 1 and is what most consumer-GPU workflows load.

The launch demo that made the rounds — a 10-second 720p image-to-video clip generated in 6.8 seconds — ran on two NVIDIA GB200 chips, so treat it as a ceiling demonstration, not an expectation. What matters for home labs is that Lightricks and ComfyUI did the quantization work themselves: the official template loads a vendor-produced int8 transformer rather than the ~42GB bf16 checkpoint.

Capability-wise, the official spec covers text-to-video, image-to-video, and first-frame/last-frame conditioning, up to 4K at up to 50 fps with synchronized audio, clips up to 20 seconds — with real limits attached: anything longer than 10 seconds drops you to 720p or 1080p at 24/25 fps, and 48/50 fps output caps at 10 seconds. The new tricks are native multi-shot generation (wide, medium, and close-up cuts of the same scene in one generation, holding character and voice consistent across cuts) and auto duration, which predicts clip length from the described action.

The license is the usual modern-open-weights shape: free to use, including commercially, for organizations under $10M annual revenue; bigger shops negotiate. Same family of terms as FLUX and friends — fine for every home lab and nearly every indie studio.

The VRAM math: what 22B of video model actually occupies

The bf16 transformer alone is roughly 42GB — irrelevant to consumer cards. The number that matters is the int8-convrot build. On the 2.3-series (the same 22B DiT architecture), the transformer-only int8 file weighs ~21.5GB for both dev and distilled; the 2.5 int8 transformer is the same class of object, and measurement backs that up.

The most useful public measurement so far is a self-hosted recipe running LTX-2.5 on an RTX 4090 via ComfyUI’s official int8 template:

• Transformer resident: 20.03 GiB
• All weights resident during sampling and decode: 22.67 GiB — on a card with 24 GiB
• 5-second clips complete normally; 10-second clips can stall right before the video appears, because the VAE decode step is where the last GiB runs out

That last line is the single most useful thing to know before you queue your first long clip. The generation isn’t failing — the decode is. If your 24GB card hangs at 100% with no output on a 10-second job, that’s not a broken install; drop to 5 seconds, or launch ComfyUI with offloading enabled (--lowvram) and accept the speed penalty on the decode step. This is the same sysmem-spillover behavior we’ve mapped for LLMs in the shared-GPU-memory fix guide — video gen just hits it at the very end of the pipeline, which is crueler.

The full ComfyUI file set, per the official ComfyUI tutorial, is bigger than the transformer alone — LTX-2.5 ships with a custom Gemma 4 12B text encoder:

ComfyUI/models/
├── diffusion_models/
│   └── ltx-2.5-22b-distilled-transformer-comfy-int8-convrot.safetensors
├── text_encoders/
│   ├── gemma4-12b-with-proj-ltx-2.5-comfy-int8-convrot.safetensors
│   └── gemma4_e2b_it_bf16.safetensors
└── vae/
    ├── ltx-2.5-video-vae-bf16.safetensors
    └── ltx-2.5-audio-vae-bf16.safetensors

Budget 35GB+ of disk for the set and be realistic about system RAM: community low-VRAM reports consistently show the model is RAM-hungry when anything offloads — 32GB is the working floor, and the most extreme low-VRAM report (6GB card, 720p 10-second clips) leaned on 44GB of system RAM. If you’re still on 32GB of DDR5, our system-RAM guide was written for LLMs but the buying logic transfers unchanged.

Note the int8-convrot format choice is quietly friendly to older cards: it uses INT8 tensor cores, which every NVIDIA card since Turing has, rather than FP8 math, which requires Ada or newer. That’s why a 2020 Ampere card is a first-class citizen here instead of falling off the supported list.

Real generation times, card by card

Numbers below are from NVIDIA’s official guide, vendor docs, and community reports — labeled accordingly. There is no independent benchmark suite for this model yet (it’s been out eight days), so ranges beat false precision.

GPU	VRAM	720p/24 clip	Time	Source type
2× GB200	384GB	10s	6.8s	vendor headline
RTX 5090	32GB	4s	~25s	NVIDIA guide
RTX 5090	32GB	8s	~3 min (weight streaming)	NVIDIA guide
RTX 4090	24GB	5s	completes normally, int8	community measurement
RTX 3090	24GB	5s	~2.5 min (distilled, no SageAttention)	community report
6GB card + 44GB RAM	6GB	10s	works, heavily offloaded	community report

The RTX 5090 numbers come from NVIDIA’s own RTX AI video guide: a 720p, 24fps, 4-second clip in about 25 seconds — but stretch the same settings to 8 seconds and generation takes around three minutes, because the job no longer fits in 32GB and ComfyUI’s weight streaming starts shuttling layers through system RAM. Even the fastest consumer card on sale crosses its VRAM cliff on this model. That’s the clearest signal of what LTX-2.5 actually is: a datacenter-class model that quantization drags onto consumer hardware, not a lightweight model that happens to scale up.

On Ampere, an RTX 3090 owner running ComfyUI’s image-to-video template reports the distilled int8 model “works great and fast for 5sec vids” at roughly 2 minutes 30 seconds per clip without SageAttention — on a card you can buy used for $1,201–$1,299. For calibration: reading these numbers against the older Wan-generation landscape, where full-quality tiers started at 80GB, a four-year-old card producing audio-synced 720p video in 2.5 minutes is the actual news here, more than the GB200 demo.

Comparative context from the 2026 local-video shootouts: community comparisons place the LTX-2 family at 3–4× the generation speed of Wan 2.2 and HunyuanVideo, and it’s the only current open video family that runs on 12GB cards at all. Wan still wins on photorealistic human detail — faces, skin, hair — and Hunyuan on physics-heavy motion, but both output silent video and both want far more VRAM for their best tiers. If your bar is “watchable footage with sound, generated on my own hardware, this decade,” LTX-2.5 is currently the only game at consumer prices.

Setup: the honest version

The install itself is the easy part — update ComfyUI, open the template browser, pick one of the three LTX-2.5 workflows (text-to-video, image-to-video, or first/last frame), and let it fetch the models. ComfyUI shipped the templates on launch day, and the distilled defaults (8 steps, CFG 1) are already wired in.

The failure modes are the usual ComfyUI suspects, and the fixes are on this site already:

• Stall at the very end of generation on 24GB → the VAE decode OOM described above. Shorter clip, or --lowvram.
• Torch reinstalled itself as CPU-only after a custom node update → the cu128 reinstall procedure.
• Black or corrupt output frames → same family of VAE-precision issues as the black-image fix guide; the bf16 VAEs here are less prone than SDXL’s fp16 was, but the diagnosis path is identical.
• Windows page-file errors while 35GB of weights load → the os error 1455 fix.

One genuinely new setting deserves attention: multi-shot. Prompt it like a shot list (“wide shot of the kitchen; cut to medium shot; close-up on her hands”) rather than one run-on scene description — the model holds character, lighting, and voice across the cuts. This replaces the janky workflow of generating separate clips and praying the character survives the seam, which was the single biggest practical complaint with the MiniMax H3 ComfyUI pipeline.

Which GPU tier should actually run this

32GB — RTX 5090: the comfortable tier. Everything short of 8-second-plus clips runs weight-resident, and ~25 seconds per 4-second clip is fast enough for iterative creative work rather than batch-and-wait. If you were already eyeing the 5090 for LLM work, video capability now stacks on top of that math. What it doesn’t do is escape the VRAM cliff entirely — even the 5090 streams weights on longer clips.

24GB — used RTX 3090 / RTX 4090: the value tier. This is the tier the release actually changed. The int8 distilled build fits with about 1.3 GiB to spare, 5-second clips work today, and the used RTX 3090 at ~$1,250 remains the cheapest 24GB ticket — 2.5 minutes per clip is batch-workflow speed, not interactive speed, but it’s local, private, and audio-synced. An RTX 4090 does the same job meaningfully faster with FP8 paths available to it.

16GB — RTX 5060 Ti / 4060 Ti / 5080: the patience tier. The int8 transformer alone (~21.5GB) exceeds your VRAM, so every generation runs partially offloaded through weight streaming. It works — that’s new, and it’s why the low-VRAM YouTube tutorials exist — but generation times balloon and 32GB+ of system RAM stops being optional. Try it on the card you own; don’t buy a 16GB card for this.

8–12GB: proof-of-concept only. The 6GB-card/44GB-RAM report proves the offload path functions. A clip that takes most of your coffee break is a demo, not a workflow. This tier is better served renting: an RTX 5090 on RunPod runs $0.99/hour — at ~25 seconds per 4-second clip, a dollar buys you an evening of testing whether LTX-2.5’s output quality clears your bar before any hardware money moves. That’s the right order of operations anyway.

If this article is the thing that finally pushes you to a bigger card, the GPU buying guide covers the full $300–$3,000 decision tree — and note that video generation rewards VRAM capacity even harder than LLM inference does, which shifts the guide’s usual bandwidth-first logic one notch toward capacity. For the open-source stack around models like this, our sister site aifoss.dev tracks the serving-side tooling in more depth.

FAQ

Can my 24GB card really run a 22B video model? Yes — via the vendor’s int8 quantization, which ComfyUI’s official template loads by default. Measured usage is 22.67 GiB all-in on an RTX 4090 during sampling and decode. Keep clips at 5 seconds and close VRAM-hungry apps first; the decode step has almost no headroom.

Does it need an RTX 50-series for the quantized build? No — this is the pleasant surprise. The int8-convrot format runs on INT8 tensor cores (Turing and newer), not FP8 units, so RTX 30-series cards run the same checkpoint. A 3090 does ~2.5-minute 5-second clips on the distilled model per community reports.

Is the audio actually generated, or added after? Generated — video and synchronized audio come out of one forward pass through the model, decoded by a separate audio VAE. It’s the first open-weights model family to do this; Wan and Hunyuan both output silent video.

Distilled or Dev — which do I download? Distilled, unless you have VRAM to burn. It samples in 8 steps at CFG 1, which is where the consumer-card generation times in this article come from. Dev is the quality-ceiling checkpoint for people running 32GB+ or cloud hardware.

Is it free for commercial use? Under $10M annual revenue, yes — the weights are open and the license covers commercial output. Above that, Lightricks wants a conversation.

How does it compare to Wan for quality? Wan 2.2+ still renders more photorealistic humans, and Hunyuan handles complex physics better. LTX-2.5 wins on speed (3–4× per community comparisons), VRAM reach, native audio, and multi-shot continuity. For most home-lab use — storyboards, social clips, prototyping — the trade favors LTX; for a hero shot of a human face, it doesn’t yet.

Recommended Gear

• Used RTX 3090 24GB — cheapest card that runs the int8 build weight-resident (~$1,250 used)
• RTX 4090 24GB — same 24GB fit, faster silicon and FP8 support
• RTX 5090 32GB — the interactive-speed tier: ~25s per 4-second clip

Sources

• LTX-2.5: Lightricks Launches 22B Open Video Model — ComfyUI Wiki
• LTX-2.5 can generate a 10-second AI video from an image in just 6.8 seconds on Nvidia superchips — VentureBeat
• LTX-2.5 — Lightricks model card, Hugging Face
• LTX-2.5 model documentation — LTX Docs
• LTX-2.5: ComfyUI Workflow Examples — ComfyUI Docs
• LTX-2.5 on RTX 4090: Fitting the 22B Audio-Video DiT in 24 GB via ComfyUI int8 — Smeltcore
• Quick Start Guide For LTX-2 In ComfyUI — NVIDIA GeForce News
• LTX-2.5 Comes to ComfyUI: Native Multi-Shot AI Video Generation — Kombitz
• RTX 3090 image-to-video template report — club-3090 GitHub discussion #481
• LTX 2.3 distilled int8-convrot transformer: size and VRAM — LTX Workflow
• Low-VRAM LTX workflow discussion (6GB card, 44GB RAM) — Kijai/LTXV2_comfy, Hugging Face
• Local AI Video Generation: Wan 2.2 vs LTX-2 vs HunyuanVideo — Local AI Master
• Best Local AI Video: Wan vs LTX & More — InsiderLLM
• GPU Cloud Pricing — RunPod
• RTX 3090 Used GPU Price & Fair Asking Range — ResalePrices
• RTX 5090 Price Tracker, August 2026 — VideoCardPrices

Last updated August 19, 2026. Prices, model availability, and VRAM behavior change fast — verify current listings and the latest ComfyUI template before buying or downloading.