LTX-2.5 (22B, Video + Audio)
Generate video with synchronized audio using LTX-2.5 on a Clore.ai RTX 5090 — 22B open-weight DiT with day-one ComfyUI support
Most open video models make you assemble the soundtrack separately. LTX-2.5 generates picture and audio together, from text, from a still image, or from an existing clip. It is also the first open video model whose official hardware guidance — 32GB of VRAM — maps exactly onto a card the Clore.ai marketplace has in volume: there were 121 free RTX 5090s at the last snapshot, at $0.25–0.77 per GPU hour.
Key Specs
Parameters
22B (asymmetric dual-stream DiT)
Inputs
Text, image, video
Output
Video with synchronized audio
Resolution
720p → native 4K
Clip length
6–20 seconds
License
LTX-2.x Community License (free commercial use under $10M revenue)
Release Date
11 August 2026
Tooling
ComfyUI (native), LTX API, diffusers
Requirements
Full precision
~66GB
1× RTX PRO 6000 96GB — $0.92–1.38/hr
FP8 (official guidance)
~32GB
1× RTX 5090 32GB — $0.25–0.77/hr
int8 / FP8 with CPU offload
16–24GB
1× RTX 4090 or 3090 — $0.07–0.42/hr, slower
Also budget 100GB of free disk for the checkpoints, and prefer a server with a fast link: the download is the slowest part of the first run.
Blackwell needs a current CUDA stack. On RTX 50-series, use a CUDA 12.8+ image — yanwk/comfyui-boot:cu130-slim for ComfyUI, or pytorch/pytorch:2.11.0-cuda12.8-cudnn9-devel if you are building your own. A CUDA 12.4 image will fail on these cards. See CUDA & PyTorch Compatibility.
Deploy on Clore.ai with ComfyUI
Rent an RTX 5090 from the Marketplace with at least 100GB of free disk.
Deploy the ComfyUI image:
Accept the model licence on the LTX-2.5 model page and create a Hugging Face token — the repo is gated.
Pull the weights and the LTX ComfyUI nodes:
Restart ComfyUI and load the official LTX workflow template — a two-stage graph that takes an image plus a text prompt. Building the node graph by hand is not worth it; start from the reference workflow and modify.
Clore.ai GPU Recommendations
1× RTX 5090
32GB
FP8, 720p–1080p
$0.25–0.77/hr
1× RTX 4090
24GB
int8 + CPU offload
$0.14–0.42/hr
1× RTX 3090
24GB
int8 + CPU offload, slowest
$0.07–0.21/hr
1× RTX PRO 6000
96GB
full precision, 4K
$0.92–1.38/hr
2× RTX 5090
64GB
full precision, faster batches
$0.50–1.54/hr
A 10-second clip costs cents. At $0.29/hr median for an RTX 5090, even a slow generation of a few minutes lands under two cents of GPU time. Rent, generate a batch, close the order.
Use Cases
Video with sound in one pass — no separate audio model, no manual sync
Image to video — animate a still with a text prompt driving the motion
Video to video — restyle or extend existing footage
4K output — native, not upscaled, on a 96GB card
Short-form content pipelines — 6–20 second clips is exactly social-format length
Previz and storyboards — cheap iteration before committing to a real shoot
Troubleshooting
403 downloading weights
The repo is gated — accept the licence on the model page, then pass --token
no kernel image on RTX 5090
CUDA 12.4 image; switch to cu130-slim or a cu128 PyTorch base
OOM on 24GB
Enable FP8/int8 and CPU offload in the workflow, drop resolution to 720p, shorten the clip
Out of disk mid-download
Needs ~100GB free; filter the marketplace for larger disks
Audio out of sync
Use the official workflow rather than a hand-built graph — the two-stage template handles the audio branch
Very slow first run
Model load and compile; the second generation is much faster on the same order
Next Steps
Alternatives: Wan 2.2 VBVR · HunyuanVideo · FramePack (6GB VRAM)
Comparison: Video generation models compared
Post-processing: RIFE interpolation · Real-ESRGAN upscaling
Links
LTX-2.5 on Hugging Face (gated)
Rent a GPU: RTX 5090 · RTX 4090 · Marketplace
Last updated
Was this helpful?