LTX-2.5 Gemma-4 12B NVFP4 for ComfyUI

ComfyUI-native NVFP4 quantization of the Gemma-4 12B text encoder used by Lightricks/LTX-2.5.

Model file

gemma4-12b-with-proj-ltx-2.5-nvfp4.safetensors

About

This checkpoint was quantized from the official LTX-2.5 BF16 text encoder:

gemma4-12b-with-proj-ltx-2.5-bf16.safetensors

It is intended as a drop-in replacement for the LTX-2.5 Gemma text encoder in ComfyUI.

The checkpoint uses a mixed-precision layout: the heavy Gemma decoder Linear weights are stored as Comfy-native NVFP4 while embeddings, normalization layers, vision components, and LTX-specific projection layers remain at their source precision.

The model was not retrained or fine-tuned.

Quantization

Quantized using Comfy-Org/comfy-quants.

The following decoder matrices were converted to NVFP4:

  • self_attn.q_proj
  • self_attn.k_proj
  • self_attn.v_proj on sliding-attention layers
  • self_attn.o_proj
  • mlp.gate_proj
  • mlp.up_proj
  • mlp.down_proj

Gemma-4 12B has 48 decoder layers with eight global-attention layers:

5, 11, 17, 23, 29, 35, 41, 47

Those global-attention layers reuse K as V and therefore do not contain a separate v_proj.weight.

Total NVFP4 matrices: 328.

ComfyUI usage

Place the file in:

ComfyUI/models/text_encoders/

Then select:

gemma4-12b-with-proj-ltx-2.5-nvfp4.safetensors

in the same LTX-2.5 text encoder loader you normally use.

No other workflow changes should be required.

Results

Tested successfully with LTX-2.5 in ComfyUI.

In testing, the NVFP4 encoder:

  • loaded and ran successfully
  • reduced VRAM usage
  • produced good output quality with no obvious degradation in the tested generations

These are practical observations rather than a formal benchmark.

Compatibility

Native NVFP4 execution is intended for compatible NVIDIA Blackwell hardware with an up-to-date ComfyUI / comfy-kitchen stack.

Behavior on other hardware may differ depending on the available fallback path.

License

This is a community-made quantized derivative of LTX-2.5 and is subject to the LTX-2.x Community License Agreement.

This is not an official Lightricks or Comfy-Org release.

Downloads last month
13
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for Deadshot699/ltx-2.5-gemma4-12b-comfy-nvfp4

Quantized
(23)
this model