LTX-2.5 generates video and synchronized audio in a single pass — a dual-stream DiT with a 4096-wide video path, a 2048-wide audio path, and cross-modal attention joining them. The bf16 transformer is 39 GB. These quants bring it to 8-22 GB.

All original licensing terms and usage restrictions carry over from the base model.

The config metadata problem

If you convert LTX-2.5 to GGUF yourself, it will not load. These files have the fix baked in; this section explains what it is, because it isn't obvious and it cost a night to find.

ComfyUI derives most model dimensions from tensor shapes. For LTX-2.5 it can't: the connector widths, the audio path dimension, and the 9-row modulation tables are only knowable from a config blob stored in the original safetensors __metadata__. ComfyUI reads it here:

dit_config.update(json.loads(metadata["config"]).get("transformer", {}))

GGUF has no __metadata__ field, so every standard conversion drops it. ComfyUI then builds an LTX-2.3-shaped model and pours 2.5 weights into it, producing a wall of shape mismatches:

size mismatch for audio_embeddings_connector.learnable_registers:
  copying a param with shape torch.Size([128, 2048]) from checkpoint,
  the shape in current model is torch.Size([128, 3840])
size mismatch for transformer_blocks.0.scale_shift_table:
  shape torch.Size([9, 4096]) from checkpoint, current model is torch.Size([6, 4096])

These files carry the full 61-key transformer config as a GGUF KV field named config. ComfyUI-GGUF surfaces every scalar KV as metadata, ComfyUI merges it, and the model builds with the right dimensions. No custom node, no config pasting — the stock Unet Loader (GGUF) just works.

Files

Quant	Size	Notes
Q8_0	~22 GB	Near-lossless reference
Q6_K	~18 GB	Excellent
Q5_K_M	~15 GB	Very good
Q4_K_M	~13 GB	Recommended
Q4_K_S	~13 GB	Slightly smaller
Q3_K_M	~11 GB	Usable, softer detail
Q2_K	~9 GB	Smallest; quality drops sharply

Grab one file rather than the repo:

hf download realrebelai/LTX-2.5_GGUFs LTX-2.5-Distilled-Q4_K_M.gguf --local-dir ComfyUI/models/unet

Companion files

The GGUF replaces only the transformer. You also need:

File	Folder
`gemma4-12b-with-proj-ltx-2.5` text encoder	`models/text_encoders`
LTX-2.5 video VAE	`models/vae`
LTX-2.5 audio VAE	`models/vae`

Two VAEs. Missing the audio one is the most common reason a render comes out silent.

The text encoder is not stock Gemma — it carries LTX's own projections (text_embedding_projection, audio_projector) that map prompt embeddings into the DiT's conditioning. A plain Gemma checkpoint will not substitute.

Settings

Distilled — few-step schedule, CFG 1.0. Guidance is baked into the weights; applying it again degrades output regardless of quantization.

Dev — full step count with real CFG.

Use ComfyUI's built-in LTX-2.5 template and swap its model loader for Unet Loader (GGUF). Everything else stays as the template ships it.

Conversion notes

Converted from the official bf16 transformer with a streaming converter, quantized with llama-quantize, then repaired.

Kept at high precision rather than quantized:

patchify_proj, audio_patchify_proj, to_gate_logits, scale_shift_table,
keyframes_abs_pos_embedding, caption_projection, time_embed, proj_out, rope

to_gate_logits matters more than its size suggests. Every attention module has one, and they gate how much signal crosses between the audio and video streams. Quantize them and video still looks fine while audio desyncs or degrades — a failure that's easy to ship and hard to diagnose.

llama-quantize does not preserve these layers on its own; they are restored afterward and verified per file. keyframes_abs_pos_embedding also has its original [1, 4096] shape recorded, since ggml trims trailing singleton dimensions on rewrite.

Credits

Lightricks — LTX-2.5
city96 — ComfyUI-GGUF
ComfyUI

Quantized by RealRebelAI · GitHub · X