BennyDaBall/LTX-2.5-22b-distilled-nvfp4-comfy-v2

LTX-2.5 22B Distilled — NVFP4 ComfyUI-ready, V2 (keyframes edition)

Caught 'em lacking. On a Sunday.

The updated official NVFP4 export (2026-08-16, "Restore keyframes support") — stamped so ComfyUI actually loads it. Again.

Text-to-video with this checkpoint — 1920×1088, native audio, 70s on an RTX 5090 (cold load included).

The saga so far

• Aug 11 — LTX-2.5 drops. The official NVFP4 transformer won't load in ComfyUI: the packed FP4 weights get read as dense bytes and you eat a mat2 shape error. I stamp per-layer .comfy_quant markers onto it and publish V1 — the first working comfy NVFP4 on the Hub.
• Aug 12 — the official file quietly grows the exact same markers (discussion #16 👀). Respect. They also notice the export dropped the keyframe positional embedding and pull the config flag.
• Aug 16, 12:19 UTC, a Sunday — "Restore keyframes support": the official file gets replaced with a keyframes-capable re-export. Fresh quantization, embedding restored... and the ComfyUI markers are gone AGAIN. Launch-day crash, the sequel.
• Aug 16, a few hours later, still Sunday — caught 'em lacking on the Lord's day. This repo: the new payload, stamped, render-verified, keyframes and all. Y'all have got to start test-loading these in Comfy before you hit upload. 🙏

What this is

The current official ltx-2.5-22b-distilled-transformer-nvfp4.safetensors (keyframes-capable 2026-08-16 re-export) with ComfyUI's per-layer {"format": "nvfp4"} .comfy_quant markers added. Weights, scales, and license metadata untouched. 7,877 tensors, 1,176 quantized layers stamped — and the headline vs V1: keyframes_abs_pos_embedding is back in the file.

Verified before upload (RTX 5090, ComfyUI 0.33)

First/last-keyframe (flf2v) bridge rendered with this exact file — two keyframes in, the model fills the 5 seconds between:

First and last frames supplied as keyframes — everything between is generated. Identity, costume and set hold, and it lands the end pose.

• flf2v: 5s @ 1280×736 in ~55s wall with SageAttention 3.
• t2v / i2v load and render clean in the standard LTX-2.5 template graphs (plain UNETLoader).

Same seed, same prompt, both quants

NVFP4 V2 — this repo (18.7GB)	Official comfy-int8-convrot (21.5GB)

10 seconds @ 1280×736, seed 910090 on both. Similar and close, not identical — and the 4-bit file is ~3GB lighter.

V1 or V2?

This one. V1 was built from the pre-keyframes export — it still works for t2v/i2v, but it physically does not contain the keyframe embedding. V2 is the current official payload, so new pulls should come here.

Usage

Drop the file in models/diffusion_models/, select it in UNETLoader in any LTX-2.5 template graph. Example t2v API workflow included in the repo.