LTX 2.5 22B Distilled — Hybrid W4A8 / INT8 ConvRot for ComfyUI

Community quantization pack of the official LTX 2.5 22B Distilled transformer and Gemma 4 12B text encoder, prepared by Agustín Meriles (ameriles) for ComfyUI.

The transformer uses a mixed recipe designed to retain audio/lip-sync behavior and visually sensitive projections while keeping most of the speed and memory advantages of W4A8:

• 900 Linear layers: INT8 tensorwise + ConvRot
• 540 Linear layers: asymmetric W4A8
• all other tensors: preserved from the official BF16 source

The text encoder uses asymmetric W4A8 for 328 tensors and preserves 358 tensors.

Files

The VAE, Audio VAE and latent upscalers are not redistributed here. Download them from the official Lightricks/LTX-2.5 repository.

Hybrid quantization recipe

The recipe keeps the audio, cross-modal and visually sensitive projections in INT8 ConvRot while using W4A8 for the remaining quantized Linear layers. The visually sensitive families are:

• video blocks: attn1.to_v, attn2.to_v, ff.net.2
• video connector: attn1.to_v, ff.net.2

The exact selector, strict expected layer counts and tests are published in ameriles/comfy-quant-bench.

Compatibility

• ComfyUI-only prequantized Safetensors format
• tested with ComfyUI v0.37.0 (73c9bad)
• tested with PyTorch 2.9.1+rocm6.4
• converted with Comfy Kitchen 0.2.35
• validated on 3 × AMD Radeon RX 6700 XT (RDNA2)

The formats are not exclusive to RDNA2. Comfy Kitchen provides multiple runtime backends, but performance and kernel availability depend on the installed ComfyUI, PyTorch, backend and GPU. This pack has only been performance-tested on the setup above.

Important text-encoder note

The W4A8 text encoder materially reduces file size and model storage. In stock ComfyUI, text-encoder execution may dequantize the weights because the CLIP path requests full-precision matrix multiplication and forced weight casting. Therefore, treat the text encoder as a memory/storage optimization first, not as a guaranteed native W4A8 speed-up.

Workflows

Two cleaned workflows are included:

• LTX2.5_Hybrid_W4A8_INT8_MultiGPU.json: the measured 3-GPU layout, using ComfyUI-MultiGPU plus the included Free VRAM and Auto Tiled VAE helper nodes.
• LTX2.5_Hybrid_W4A8_INT8_Simple.json: standard ComfyUI loaders and tiled VAE decode, with no MultiGPU or memory-helper dependency.

The Simple workflow is portable, but it does not imply that the full pipeline will fit entirely in a 12 GB GPU. ComfyUI may rely on CPU offload, and speed depends heavily on VRAM and system RAM.

The MultiGPU reference allocation is:

• transformer: cuda:1,34%;cuda:2,66%
• text encoder: cuda:0,50%;cpu,50%
• video VAE: cuda:0

Setup

1. Install/update ComfyUI with native quantized Safetensors support and Comfy Kitchen.
2. Place the transformer and text encoder in the directories shown above.
3. Download these official LTX 2.5 components as required by the workflow:
   • ltx-2.5-video-vae-conv-bf16.safetensors
   • ltx-2.5-audio-vae-bf16.safetensors
   • ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors
   • ltx-2.5-latent-temporal-upscaler-x2-bf16-1.0.safetensors
4. For the MultiGPU workflow, install ComfyUI-MultiGPU and copy the included custom_nodes/ComfyUI-LTX25-Hybrid-Helpers folder into ComfyUI/custom_nodes/.
5. Load one of the workflows, select your own image/audio inputs and adjust device allocations for your hardware.

Measured result on the reference setup

For the tested 5-second, two-stage talking-head workflow, four warm runs completed in 461.16–465.90 seconds (about 7:43 average). Earlier Q4_K_M runs on the same evolving workflow averaged about 13:13. This is a setup-specific observation, not a universal benchmark: workflow revisions, thermal state, PCIe topology, prompts and input media all affect end-to-end time.

Qualitative validation covered prompt following, Spanish speech, lip sync, teeth/facial detail and repeated 5-second generations.

Example generations

The following 5-second samples were generated with the quantized transformer and text encoder from this repository. They cover English and Spanish dialogue, lip sync, facial expression, camera movement and image-conditioned generation.

Quantization provenance

Source files from Lightricks/LTX-2.5:

• diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors
• text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors

Transformer conversion:

• preset: ltx25-q4km-audio-balanced-visual-sensitive
• output: 17,045,068,544 bytes
• INT8 vs BF16 sampled relative-error median: 0.0094466
• INT8 vs BF16 sampled relative-error maximum: 0.0110030

Text-encoder conversion:

• format: asym_w4a8_int8
• group size: 16
• ConvRot group size: 256
• output: 10,604,318,782 bytes

See the JSON sidecars for machine-readable details and SHA256SUMS for integrity hashes.

Compatibility notes

• ComfyUI-only; these files are not intended for the standard ltx-pipelines PyTorch loader.
• Performance is validated on ROCm/RDNA2 only.
• Quality validation was performed with talking-head, dialogue and image-conditioned samples rather than a formal perceptual benchmark.
• Five-second generation is the validated reference workload; longer clips require additional VRAM or a different memory allocation.

License and modification notice

This is a quantized derivative of LTX 2.5 and is distributed under the LTX-2.x Community License Agreement. Read the complete included LICENSE-LTX-2.x, including its commercial-use conditions and use restrictions, before downloading or using the files.

The original BF16 files were modified by quantization in October 2026 by Agustín Meriles. Lightricks did not produce or endorse this quantization.

Credits

• Base model: Lightricks/LTX-2.5
• Quantization tooling and research base: JoaoZaokk/comfy-quant-bench
• Quantization, testing and release: Agustín Meriles / ameriles