You need to agree to share your Hugging Face information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

These weights are a derivative of LTX-2.5 under the LTX-2.x Community License Agreement. An entity with annual revenue of US$10 million or more (counted with its affiliates) must hold a paid license from Lightricks before any use other than the non-commercial uses in Section 2.2, and these weights may not be transferred to such an entity without that license (Section 3.5). Use is subject to Attachment A and the Lightricks Acceptable Use Policy.

Log in or Sign Up to review the conditions and access this model content.

LTX-2.5 FP8 for Paiton on RDNA4

These are weights of Lightricks/LTX-2.5 (22B distilled, video with audio) converted for one AMD Radeon AI PRO R9700 (gfx1201, 32 GB). The transformer and the Gemma 4 text encoder are in FP8 in layouts that Paiton's native HIP kernels read. The files only run with the Paiton container; they do not load in diffusers, ComfyUI or Lightricks' ltx-pipelines.

Speed

Request (121 frames, 5 s at 24 fps) Paiton, this repository Stock LTX-2.5, BF16 Speedup
text to video, 1280 x 704 18.1 s 128.6 s 7.1x
image to video, 1280 x 704 21.1 s - -
text to video, 1920 x 1088 48.8 s 420.5 s 8.6x
image to video, 1920 x 1088 51.9 s - -

One R9700 at its 300 W limit. Paiton: server time from HTTP request to mp4 with a warm server (median of 3 runs), peak VRAM use on the card 26.6 GiB at 1280 x 704 and 27.4 GiB at 1920 x 1088. Stock: Lightricks' ltx-pipelines distilled pipeline in BF16 on the same card, warm, with --offload cpu because the BF16 weights do not fit in 32 GB. With --quantization fp8-cast the stock pipeline is slower on this card (156.7 s and 488.5 s): its matrix products still run in BF16.

Run

git clone --depth 1 https://github.com/Eliovp-BV/paiton-vllm-plugin.git
cd paiton-vllm-plugin
export HF_TOKEN=<your Hugging Face read token>
./models/LTX-2.5/launch.sh
./models/LTX-2.5/run.sh generate "A street drummer plays a fast rhythm on a city sidewalk at dusk while the crowd claps and cheers" --out drummer.mp4

Accept the license terms on this page first and use a token of the same account. The first start downloads this repository (38 GB) at the revision pinned in the container and checks every file against its SHA-256. The server answers POST /v1/videos/generations on port 30000 with an mp4 (H.264 video, AAC audio). The guide lists the request fields and limits.

Quality

Any quantized version of this 11-step sampler, the stock pipeline's FP8 cast mode included, follows a different path from the BF16 one after small numeric changes, so its frames differ from the BF16 video in layout and detail. PSNR and SSIM against a BF16 run with the same noise therefore measure how far the sampler drifted from that run, not how good the video looks. With the current FP8 weights (round to nearest):

Case, 121 frames This build, PSNR / SSIM Stock --quantization fp8-cast, PSNR / SSIM
text to video, 1280 x 704 25.7 dB / 0.918 25.3 dB / 0.90
image to video, 1280 x 704 24.5 dB / 0.881 -
text to video, 1920 x 1088 18.9 dB / 0.724 18.7 dB / 0.726
4 more prompts, 1920 x 1088 18.0 to 25.5 dB 19.4 to 26.4 dB
mean of the 5 prompts at 1920 x 1088 21.25 dB / 0.726 22.35 dB / 0.759
audio of the same 5 prompts, mean log-spectral distance (lower is closer) 4.69 dB 4.75 dB

At 1280 x 704 this build is as close to the BF16 run as the stock FP8 cast mode. At 1920 x 1088 it drifts about as much: 1.1 dB lower on the mean of five prompts, with two prompts lower (by 1.9 and 4.1 dB) and three equal or higher. Side by side, the frames of the two FP8 runs show different camera paths and positions, with no blocking, banding or noise. We also tried GPTQ rounding for the transformer weights; the videos showed no consistent gain, so this release keeps round to nearest.

Files

Folder Content
config.json Model name, base model, license and the container tag; the container fetches it once per install
dit/ Transformer: FP8 block weights (dit8.bin), the BF16 tensors outside the blocks (dit_rest.bin), precomputed AdaLN modulation for the 9 sigma values of the distilled schedule (tables.bin)
te/ Gemma 4 12B text encoder, feature extractor and embedding connectors in FP8, as 65 streaming units; tokenizer
vae/ Unchanged copies of the video VAE, the x2 latent upscaler and the audio VAE with vocoder

MODIFICATIONS.md describes every change to the original files.

License

LTX-2.5 and these weights are under the LTX-2.x Community License Agreement (license date 11 August 2026, amended 29 October 2026). The full text is in LICENSE; it governs, and the summary below does not replace it.

  • These weights are a Derivative of LTX-2.x (Section 1.5). We distribute them under that agreement only, with a full copy of it, Lightricks' notices and a notice of our changes (Sections 3.2 to 3.4).

  • Entities with annual revenue of US$10 million or more, counted with their affiliates, need a paid commercial license from Lightricks (licensing@ltx.io) for any use other than the non-commercial uses in Section 2.2 (Section 2.1). Such an entity must hold that license before it receives or uses these weights (Section 3.5).

  • Every use, every output and every further derivative is subject to Section 4, Attachment A and Lightricks' Acceptable Use Policy. If you pass the weights on or let others use them, your terms must include those restrictions and you must tell the recipients they apply (Sections 3.1 and 4).

  • State clearly that generated content is machine generated when you publish it (Attachment A, item 5).

  • Do not remove watermarks, provenance data or other disclosure features (Section 6). The Lightricks reference pipeline adds a watermark only when given a marking pack (--marker-root-path, off by default); the Paiton runtime does not add one.

  • Attachment A, item 20 excludes products or services that compete with Lightricks' own commercial offerings unless you hold a separate commercial license.

  • The Paiton container that runs these weights is Apache-2.0 for ElioVP's code. Its native runner and libraries implement the LTX-2.x architecture, so the LTX-2.x Community License Agreement applies to them as well (Section 1.5 (ii)). The container's THIRD_PARTY_NOTICES.md lists every license in it.

Lightricks does not endorse this repository.

Downloads last month
19
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for EliovpAI/LTX-2.5-FP8-Paiton-RDNA4

Quantized
(48)
this model