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 12 artifacts passed structural and deep DDB-qanalyse validation. The three transformer formats 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.
• int8 is the precision-retaining INT8 profile. Five sensitive transformer matrices and eight text-encoder matrices remain BF16.
• int8_lean_convrot is the standard DDB lean ConvRot profile and leaves more room for the rest of the workflow.
• nvfp4 is the smallest transformer and text-encoder profile. It loads through the supported RTX 4090/Ada fallback path; native NVFP4 acceleration is a Blackwell path.
• Quantized LoRAs use a DDB descriptor/runtime. Stock ComfyUI LoRA loading is not claimed for those three files.

Choose a profile

Choose one transformer row. Dev and Distilled have identical profile layouts; download only the route you intend to use.

Profile	Dev transformer	Distilled transformer	Matching text encoder	Optional Distilled LoRA	Intended memory route
int8	20.152 GiB	20.152 GiB	14.440 GiB	4.715 GiB	Highest precision-retention policy; use ComfyUI offload on a 24 GiB GPU.
int8_lean_convrot	20.027 GiB	20.027 GiB	14.317 GiB	4.216 GiB	Standard 24 GiB DDB ConvRot route with offload.
nvfp4	12.641 GiB	12.641 GiB	10.428 GiB	2.482 GiB	Smallest complete DDB route; supported fallback on RTX 4090 and native acceleration on supported Blackwell hardware.

File size is not complete-workflow peak VRAM. Resolution, frame count, audio, attention backend, text encoder, VAEs, latent upscalers, and ComfyUI offload settings also affect memory. A complete 16/24 GiB end-to-end VRAM matrix has not yet been measured.

Quantization layout

Component/profile	Quantized matrices or pairs	BF16 retained matrices or pairs	Artifact validation
Transformer int8	1,435	5	7,219 tensors, PASS
Transformer int8_lean_convrot	1,440	0	7,229 tensors, PASS
Transformer nvfp4	1,425	15	8,624 tensors, PASS
Gemma-4 TE int8	320	8	1,326 tensors, PASS
Gemma-4 TE int8_lean_convrot	328	0	1,342 tensors, PASS
Gemma-4 TE nvfp4	310	18	1,616 tensors, PASS
Distilled LoRA int8	1,200 pairs	460 pairs	6,920 tensors, PASS
Distilled LoRA int8_lean_convrot	1,360 pairs	300 pairs	7,400 tensors, PASS
Distilled LoRA nvfp4	1,360 pairs	300 pairs	10,120 tensors, PASS

INT8 weights use row-wise scaling plus ConvRot/Hadamard rotation with group size 256. The int8 and int8_lean_convrot names describe different precision-retention policies, not different model architectures. The BF16 islands were selected from the source tensor inventory and DDB-qanalyse risk classification; they are not a substitute for the pending perceptual A/B.

Measured runtime checks

The following checks were executed against the physical release artifacts, not header-only fixtures:

Check	Hardware/runtime	Result
Dev int8 transformer recognition and quantized projection	RTX 4090, ComfyUI embedded Python	PASS (int8_tensorwise)
Dev int8_lean_convrot transformer recognition and quantized projection	RTX 4090, ComfyUI embedded Python	PASS (int8_tensorwise)
Dev nvfp4 transformer recognition and quantized projection	RTX 4090, ComfyUI embedded Python	PASS (nvfp4)
Gemma-4 nvfp4 tokenize, text forward, and LTX video/audio projections	RTX 4090, Torch 2.13/cu130, ComfyUI ltx25 branch	PASS (LTXAVTEModel_, [1,25,6144], finite; 12.813 GB peak allocation)
NVFP4 LoRA descriptor decode and finite tensor smoke	RTX 4090, DDB experimental LoRA runtime	PASS
Complete equal-seed video/audio quality comparison	—	PENDING

These are loader/kernel checks, not complete prompt-to-decoded-video peak-VRAM or quality measurements.

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 + matching TE.
• Fast official distilled route: Distilled transformer + matching TE, using the official 8-step / CFG 1 workflow.
• Most retained source precision: int8.
• Balanced DDB 24 GiB route: int8_lean_convrot.
• Smallest complete DDB route: nvfp4 transformer + nvfp4 TE.
• Distilled LoRA experiments: matching LoRA profile plus the compatible DDB quantized-LoRA runtime; do not load it with an ordinary stock LoRA loader.

Validation and reproducibility

Every released artifact passed:

1. exact tensor count, shape, dtype, metadata, and quantization-inventory checks;
2. deep DDB-qanalyse structural, payload, architecture, and source-integrity analysis;
3. deterministic policy and build-report verification;
4. local byte-size and SHA-256 verification.

The repository includes the exact release recipe, per-artifact job files, layer policies, build reports, modification notice, and SHA256SUMS.

• Source revision: 28dac7acdc1f78a70e98687db261a949754f8941
• Converter: Comfy convert-to-quant 1.3.1
• Converter SHA-256: b929e7d387d0baf6753c8171f482285e49649201bc7085ebfa3ac12773d350f3
• Manual seed: 20260812
• 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.
• The quantized LoRA format needs the matching DDB runtime and is currently an experimental path rather than a stock-ComfyUI compatibility claim.
• The Dev and Distilled transformer layouts were validated independently, but only the Dev files have completed the physical ComfyUI loader-format smoke at the time of this card revision.

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.