LTX2.3-22b-distilled-video-only-MXFP8

LTX-2.3 22B Distilled 1.1 — experimental video-only MXFP8

Experimental R&D derivative. This repository is not an official Lightricks release and is not yet promoted to production. End-to-end GPU validation on the locked RTX 5090 workload is bit-exact in normal and transition modes; do not infer a general speed claim from the existence of these files.

These artifacts are modified derivatives of LTX-2.3. They remove parameters that are unreachable when audio generation and both audio/video cross-attention directions are disabled. Every retained tensor payload is copied byte-for-byte from its source safetensors file: there is no deserialization, cast or re-quantization.

The complete LTX-2 Community License Agreement, including Attachment A, is in LICENSE. Use and redistribution are subject to that agreement. In particular, downstream users must receive the complete agreement and notice that these are modified files.

Files

The adjacent *.manifest.json files record source/output hashes, tensor counts and exact removed byte counts.

Modification notice and provenance

Transformer source:

• Repository: Kijai/LTX2.3_comfy
• File: diffusion_models/ltx-2.3-22b-distilled-1.1_transformer_only_mxfp8_block32.safetensors
• Source SHA-256: b7a945ff24d65ad22c6977787c2e594e74df226e35f1f9dedb64be8fdbd6ffd8
• Change: removed 4,970 audio and audio/video cross-modal tensors (7,066,735,680 payload bytes); retained 2,738 tensors (16,984,850,816 payload bytes).

Projection source:

• Repository: Kijai/LTX2.3_comfy
• File: text_encoders/ltx-2.3_text_projection_bf16.safetensors
• Source SHA-256: 911d59bb4cb7708179c9a0045ea0fe41212ecfb77aed3a02702b7c0a8274911f
• Change: removed both audio_aggregate_embed tensors (770,707,456 payload bytes); retained both video_aggregate_embed tensors.

Transition LoRA source:

• Repository: joyfox/LTX-2.3-Transition-LORA
• File: loras/ltx2.3-transition.safetensors
• Source SHA-256: ba420d6fefafced8e317e2d6ff951b312b52f534377d016b491877a00b830d33
• Change: removed 192 audio_ff tensors (62,914,560 payload bytes); retained 960 video-path tensors.

All three output files carry a prominent ltx_video_only_modified=true metadata notice. The transformer discriminator is changed to image_model=ltx23_video_only, so it fails closed instead of being silently loaded as a stock LTXV model.

Required runtime

This transformer is intentionally not a stock drop-in checkpoint. It requires the accompanying ltx23_video_only ComfyUI custom node/runtime, pinned to:

• ComfyUI 1377a2f72925ed7a5518c1900ff71c6740217b0d
• ComfyUI-KJNodes e27a505b3ba6ce42687fe00500deda103d9d6071

The custom model constructs only the unchanged video modules, keeps the AV tuple contract with an empty audio tensor for KJNodes, and preserves LTX block prefetch, compressed timesteps and blocks_replace. The projection file must be loaded with LTX23VideoOnlyTextEncoderLoader.

The validation workload is fixed at 1280×704, 129 frames, 25 fps, 8 steps, CFG 1, identical scheduler/sigmas/noise and both normal I2V and transition LoRA modes. Acceptance requires equality of all eight step latents, the final latent and all 129 decoded RGB frames before any performance result is used.

That acceptance suite now passes on RTX 5090 for both modes: all eight step latents, the final latent and all 129 decoded frames are bit-identical to the full-model baseline (torch.equal=true, max_abs_error=0). The sealed reports are state/rd/exact/results/v9-normal/quality/exact.json and state/rd/exact/results/v9-morph/quality/exact.json; each binds the runtime, server and five model identities by SHA-256. Formal timing, stability and fresh-boot validation remain separate promotion gates.

Reproduction

The artifacts were produced with the repository script scripts/ltx23_video_only_checkpoint.py in strict-known-model mode. The tool reads the safetensors header, validates the pinned source contract, writes a new header and streams only the retained byte ranges to an atomic output. It does not import Torch.

Attribution

LTX-2 and LTX-2.3 are by Lightricks Ltd. The ComfyUI-formatted MXFP8 and projection artifacts used as byte sources were published by Kijai; the transition LoRA source was published by Joyfox. This derivative is an independent optimization experiment and does not imply endorsement by those authors.