DmitryDB/LTX-2.5-ComfyUI-Quants

LTX-2.5 ComfyUI Quants by DDB

Community quantizations of the official split-weight Lightricks/LTX-2.5 release for ComfyUI. The repository contains matching Dev and Distilled diffusion transformers, the LTX-specific Gemma-4 12B text encoder with projection, and an experimental quantized Distilled LoRA.

These are community conversions, not official Lightricks or ComfyOrg releases.

Pre-release validation boundary: all 10 current artifacts passed structural and deep DDB-qanalyse validation. The four final transformer files were recognized by ComfyUI on a physical RTX 4090, the NVFP4 Gemma-4/LTX text encoder passed a real tokenize + forward + dual-projection smoke, and the quantized-LoRA runtime passed a real GPU tensor-dequantization smoke. Equal-seed prompt-to-decoded-video and audio quality A/B tests are still pending, so no perceptual-quality ranking is claimed yet.

Naming

• dev is the regular configurable LTX-2.5 route.
• distilled is the official fixed 8-step / CFG 1 route.
• Transformer int8-convrot is the quality-retaining INT8 profile: 1,435 searched INT8 matrices, ConvRot/Hadamard group size 256, and five BF16 sensitivity islands selected by DDB-qanalyse.
• Transformer nvfp4 is the compact profile: 1,425 searched NVFP4 matrices and fifteen BF16 sensitivity islands. Ada uses the supported fallback path; native NVFP4 acceleration is a Blackwell path.
• The transformer lineup now has only these two profiles. Existing text-encoder and LoRA profile names are unchanged.
• Quantized LoRAs use a DDB descriptor/runtime; stock ComfyUI LoRA loading is not claimed for them.

Choose a profile

Choose one transformer row. Dev and Distilled use the same two-profile layout; download only the route you intend to use.

Profile	Dev transformer	Distilled transformer	Matching existing TE	Optional existing Distilled LoRA	Intended route
int8-convrot	20.152 GiB	20.152 GiB	int8	int8	Highest retained transformer precision; use ComfyUI offload on a 24 GiB GPU.
nvfp4	12.641 GiB	12.641 GiB	nvfp4	nvfp4	Smallest transformer route; supported fallback on RTX 4090 and native acceleration on supported Blackwell hardware.

The retired transformer Lean profile saved only about 127.9 MiB by quantizing five matrices which the source analysis marked as sensitivity islands, so it was removed. File size is not complete-workflow peak VRAM; resolution, frame count, audio, attention, TE, VAEs, upscalers, and offload also matter.

Dev v2 searched rebuild

Both Dev and Distilled transformers were rebuilt from their exact official BF16 bytes. The v2 rebuild preserves separate Q/K/V and replaces the old fixed/simple scale path with exhaustive FP32 MSE scale selection:

• int8-convrot: 101 scale candidates from 1.00 to 0.80 per selected matrix, then stock-compatible INT8 ConvRot g256 storage;
• nvfp4: 101 scale candidates from 1.00 to 0.50 per 16-value block, stored in the stock Comfy NVFP4 wire format.

Every final transformer is bound to its source SHA-256, tensor-schema digest, policy digest, kernel digest, Qfabrik build lock, output SHA-256, and a real stock-ComfyUI load plus quantized-projection smoke on an RTX 4090.

Quantization layout

Transformer profile	Quantized matrices	BF16 retained matrices	Output tensors
int8-convrot	1,435	5	7,219
nvfp4	1,425	15	8,624

The five INT8 BF16 islands and fifteen NVFP4 islands are identical policy shapes for Dev and Distilled but are independently source-bound. This is a measured quantization/runtime statement, not a perceptual quality ranking.

Measured runtime checks

All four final physical transformer files passed:

Check	Hardware/runtime	Result
Dev int8-convrot / nvfp4 stock load and projection	RTX 4090, ComfyUI embedded Python	PASS
Distilled int8-convrot / nvfp4 stock load and projection	RTX 4090, ComfyUI embedded Python	PASS

These are exact loader/kernel checks. Complete equal-seed decoded-video/audio quality and end-to-end peak-VRAM comparisons remain pending.

Installation and required upstream components

Place the selected files in:

ComfyUI/models/diffusion_models/<one Dev or Distilled transformer>.safetensors
ComfyUI/models/text_encoders/<matching text encoder>.safetensors
ComfyUI/models/loras/<optional quantized Distilled LoRA>.safetensors

A complete workflow still needs the official BF16 auxiliary files. They are deliberately not requantized or duplicated here:

• vae/ltx-2.5-audio-vae-bf16.safetensors
• vae/ltx-2.5-video-vae-bf16.safetensors
• vae/ltx-2.5-video-vae-conv-bf16.safetensors
• the official latent_upscale_models/ files used by the chosen workflow;
• the official duration model/head when required by the workflow.

Use a current ComfyUI installation and start from the official LTX-2.5 example workflows. The quantized transformers and text encoders are ComfyUI-format weights; use in ltx-pipelines is not claimed.

Which route should I download?

• Normal configurable generation: Dev transformer.
• Fast official 8-step / CFG 1 route: Distilled transformer.
• Best retained transformer precision: int8-convrot.
• Smallest transformer route: nvfp4.

Text encoder and optional experimental LoRA selection remains a separate choice; their existing profile filenames were not renamed by this transformer cleanup.

Validation and reproducibility

Every released artifact has exact tensor/header, DDB-qanalyse, build-report, byte-size, and SHA-256 evidence. The four v2 transformers additionally include exact source and tensor-schema locks, deterministic policy and kernel digests, effective Qfabrik build locks, and stock-Comfy runtime-smoke reports.

• Transformer converters: DDB searched v2 converters in recipes/v2/tools/.
• Transformer INT8 search: FP32 per-matrix MSE grid, 101 candidates, ConvRot g256.
• Transformer NVFP4 search: FP32 per-block MSE grid, 101 candidates, block 16.
• Existing TE/LoRA conversion evidence remains under its original recipe paths.
• Official quantized files were comparison inputs only, never policy authority.

Known limits

• A prompt-to-decoded-video/audio perceptual A/B has not yet been completed.
• Full workflow VRAM and speed numbers for 16 GiB, 24 GiB, and Blackwell systems are not yet published.
• Quantized LoRAs require the matching DDB runtime and are not stock-ComfyUI compatibility claims.
• The removed transformer paths remain recoverable from repository history.

License and attribution

Use is subject to the included LTX-2.x Community License Agreement. Read LICENSE.md before using or redistributing the files. The base model is by Lightricks. This community conversion is not affiliated with or endorsed by Lightricks or ComfyOrg. Original copyright and notices remain with their owners.