PulpCut/LTX-2.5-INT8-ConvRot-safetensors

LTX-2.5 · INT8 ConvRot (h3ddle subset)

The four-file subset H3ddle loads from Lightricks/LTX-2.5, based on revision 6c7e5e5.

The text encoder and both VAEs remain byte-identical to Lightricks' release. The distilled INT8 ConvRot transformer is an exact-value layout repack: its 1,344 quantized projections are transposed from [output, input] to [input, output], all 5,885 other tensors are copied byte-for-byte, and one versioned marker tells H3ddle to select the matching Metal kernel. Nothing was retrained, merged, pruned, or requantized.

Why this mirror exists

The upstream repository is gated, so an application cannot fetch it on a user's behalf without that user first accepting terms on the website. This copy carries the same Agreement — see LICENSE, which is a complete copy — so the terms travel with the weights rather than being skipped.

Using these weights binds you to the LTX-2.x Community License Agreement, including the use-based restrictions in Section 4 and Attachment A in their entirety. Read them. They restrict what you may generate, and Section 6 separately forbids circumventing watermarking, provenance or latent-disclosure mechanisms. A commercial entity as defined in Section 2 needs a paid licence from Lightricks.

What is here, and what is not

file	bytes
diffusion_models/ltx-2.5-22b-distilled-transformer-comfy-int8-convrot.safetensors	21,504,034,388
text_encoders/gemma4-12b-with-proj-ltx-2.5-comfy-int8-convrot.safetensors	15,372,969,374
vae/ltx-2.5-video-vae-conv-bf16.safetensors	1,452,269,922
vae/ltx-2.5-audio-vae-bf16.safetensors	364,866,540

38.69 GB in total, against upstream's ~180 GB. Deliberately absent: the BF16 and NVFP4 transformers, the dev (non-distilled) transformer, the LoRA, the latent upscalers, the duration head, and the non-conv video VAE. The app runs the distilled checkpoint at eight steps and does not use the stage-2 upscale ladder, so those files would be download with nothing behind it.

For anything other than that subset, go upstream — this is not a replacement for the model card, the paper, or the full release.

Measured performance versus the regular layout

Matched H3ddle runs on a 32 GiB M1 Pro used the same prompt, seed, 512x512 canvas, five-second duration, eight passes, and 65 frames. Quality was visually identical in all three runs.

comparison	regular / baseline	optimized path	resulting gain
denoising, input-major layout only	458.0 s	401.5 s	12.3% (1.14x)
total, input-major layout only	655.7 s	598.3 s	8.8% (1.10x)
denoising, shipped layout + F32 attention	458.0 s	369.2 s	19.4% (1.24x)
total, shipped layout + F32 attention	655.7 s	565.5 s	13.8% (1.16x)

The input-major improvement comes from this checkpoint. H3ddle also enables the measured F32 attention path automatically when the video sequence is long enough and its memory guard passes; no environment variables or user setup are required. Actual gains vary with geometry, duration, memory pressure, and Mac.

The repack and exact verification are reproducible with Scripts/repack-ltx-input-major.py.

Credit

LTX-2.5 is by Lightricks. All copyright, patent, trademark and attribution notices in the original release are retained here; the model, quantization and licence are theirs. This repository adds the native-runtime subset and the exact-value transformer layout described above.