LTX-2.5 in ComfyUI on 16 GB (2026): GGUF Setup, Heretic Text Encoder and Best LoRAs

LTX-2.5 from Lightricks is the open video model most people want to run right now. It is a 22B model that makes video and synchronized audio in one pass, it has an 8-step distilled version, and since August Lightricks has released more than a dozen official LoRAs for it, from slow motion to cinemagraphs. The official repo has passed 1.9 million downloads. The problem is size: the official ComfyUI files need 32 GB of VRAM, and its Gemma 4 text encoder sometimes waters down perfectly normal prompts for action, horror or fantasy scenes. This guide fixes both. It shows the exact GGUF files that run LTX-2.5 on 16 GB (and, with compromises, 12 GB) cards, where every file goes in ComfyUI, how to swap in a Heretic text encoder that keeps your prompts intact, the best settings and the official LoRAs worth installing.

What Is LTX-2.5?

LTX-2.5 is a 22B diffusion transformer released by Lightricks in August 2026. Compared with LTX-2.3 it adds a Gemma 4 12B text encoder (with LTX’s video and audio projection layers bundled into the same file), a new video VAE, an auto-duration head that picks clip length from your prompt, and native output up to 4K at up to 50 fps. Every generation can include synced sound: footsteps, rain, music or dialogue that you describe in the prompt.

For a 16 GB card, the distilled model is the one to use. It is guidance-distilled, so you run it at CFG 1 with a fixed 8-step schedule, which also halves the work per step compared with CFG above 1. The licence is the LTX-2.x Community License: free for commercial use for companies under $10M annual revenue.

VRAM Picks

Why a 15.69 GB file works on a 16 GB card: ComfyUI loads the text encoder first, encodes your prompt, then moves the encoder to system RAM before the transformer runs. It can also stream part of the transformer from RAM when VRAM runs short. That makes system RAM the real limit: 32 GB works with GGUF encoders, 64 GB is comfortable with the INT8 encoders. If Q4_K_M runs out of memory at your resolution, step down to Q3_K_M before lowering resolution.

All Files and ComfyUI Folders

The official files are in the gated Lightricks/LTX-2.5 repo. Sign in to Hugging Face and accept the licence first, or downloads return a 401 error.

Notes:

• The audio VAE is required even if you only care about the picture. A missing audio VAE is the most common cause of silent or failed renders.
• There is a lighter ltx-2.5-video-vae-conv-bf16.safetensors VAE. It decodes faster with slightly lower quality, which helps on 12 GB cards.
• LTX-2.5 has no separate projection or connector file. That was an LTX-2.3 requirement. The projections live inside the Gemma 4 encoder file, which is why any replacement encoder must include them.
• Unet Loader (GGUF) reads both models/unet/ and models/diffusion_models/, so either folder works for the GGUF.

Not sure where something goes? See the ComfyUI model folder cheat sheet. For which quant to choose, see GGUF quantization explained.

Install ComfyUI and the Nodes

1. Update ComfyUI. LTX-2.5 support is in the core nodes, so an old install will show red “missing node” boxes. In the portable build, run update/update_comfyui.bat. In a git install:

   cd ComfyUI
   git pull
   pip install -r requirements.txt

2. Install ComfyUI-GGUF (city96) for the GGUF transformer. In ComfyUI Manager search “ComfyUI-GGUF”, or manually:

   cd ComfyUI/custom_nodes
   git clone https://github.com/city96/ComfyUI-GGUF
   pip install --upgrade gguf

   Portable build: from the folder with run_nvidia_gpu.bat, run python_embeded\python.exe -s -m pip install -r ComfyUI\custom_nodes\ComfyUI-GGUF\requirements.txt.

3. Optional: ComfyUI-LTXVideo (Lightricks/ComfyUI-LTXVideo) adds the Tiled Fusion sampler, the IC-LoRA guide nodes and the 4K/8K upscale, HDR and video-to-video example workflows in example_workflows/2.5/. Install it from Manager (search “LTXVideo”) or:

   cd ComfyUI/custom_nodes
   git clone https://github.com/Lightricks/ComfyUI-LTXVideo
   cd ComfyUI-LTXVideo
   pip install -r requirements.txt

4. Restart ComfyUI and check the console for import errors.

Step-by-Step: LTX-2.5 GGUF on 16 GB

1. Download the GGUF, encoder, both VAEs and the spatial upscaler into the folders above.
2. Start ComfyUI with a little VRAM held back for the decode: python main.py --reserve-vram 1.5 (portable: add --reserve-vram 1.5 to the line in run_nvidia_gpu.bat). More launch options: ComfyUI command-line arguments.
3. Open Template Library, Video and load LTX-2.5 I2V (or T2V).
4. Replace the template’s diffusion model loader with Unet Loader (GGUF), select LTX-2.5-Distilled-Q4_K_M.gguf and connect its MODEL output where the old loader was. Leave everything else as the template ships it.
5. In the text encoder loader, pick the INT8 encoder (or the Heretic encoder, next section). Set the video VAE and audio VAE loaders.
6. Set the first stage to 960×544 and 121 frames. The template’s x2 spatial upscaler then takes it to 1920×1088.
7. Load your start image, write a prompt describing motion, camera and sound, and run. Keep the seed fixed while you test settings.

If the second (upscale) stage runs out of memory, render stage 1 only at 960×544 and upscale afterwards. The RIFE and SeedVR2 upscale workflow works on any video, including LTX output.

Even lower VRAM: the GGUF Gemma 4 encoder from elix3r/gemma4-12b-with-proj-ltx-2.5-GGUF (Q4_K_M 8.41 GB, Q5_K_M 9.51 GB) loads with CLIPLoader (GGUF) set to type ltxv. At the time of writing it needs a small patch to ComfyUI-GGUF, which is published on that model page, so treat it as an advanced option and keep a backup of the node folder.

Swap In the Heretic Text Encoder

LTX-2.5 reads your prompt with Gemma 4 12B, an instruction-tuned chat model. Chat models carry refusal behaviour, and inside a video pipeline that shows up as quiet prompt weakening: a sword fight that comes out as a stand-off, a horror creature that turns friendly, a stunt that never happens. The DeepNeuralNerd model card describes standard Gemma as having “residual refusals that can still gut perfectly innocent prompts”.

DeepNeuralNerd’s LTX-2.5 encoder is built from llmfan46’s Heretic version of Gemma 4 12B. Heretic is an automatic tool that removes the refusal direction from a model’s weights while keeping its knowledge. The important part: the file keeps the original LTX-2.5 video and audio projection weights, so it loads in the same node.

1. Download the bf16 file (or the INT8 ConvRot file for 16 GB cards).
2. Put it in ComfyUI/models/text_encoders/.
3. Refresh ComfyUI, select it in the LTX-2.5 text encoder loader in place of gemma4-12b-with-proj-ltx-2.5-....
4. Re-run a prompt that was coming out weak, with the same seed, and compare.

The author notes the encoder is mainly about prompt fidelity: it makes the model follow what you wrote. It does not add knowledge the model never learned. Use it for action, horror, fantasy, war re-enactments, VFX tests and similar creative work, and keep everything legal: no real people without consent and no deepfakes. Our uncensored local AI video guide covers the same approach for MiniMax H3 and LTX-2.3.

Best Settings

Using the dev model instead? Add ltx-2.5-22b-distilled-lora-450-bf16.safetensors (8.9 GB) to get the same 8 steps at CFG 1, or the much smaller downranked version TheDivergentAI/LTX2.5-Turbo-Distill-8step-LoRA (rank 64 is 1.32 GB) at strength 1.0, or 0.8 if the result looks over-sharp. Never apply a distill LoRA to the already distilled checkpoint.

Best LTX-2.5 LoRAs

Lightricks publishes official LoRAs for LTX-2.5 on Hugging Face. They go in ComfyUI/models/loras/. Most are gated (accept the licence like the main repo). IC-LoRAs (“in-context”) take a reference video or image as a guide and are used with the IC-LoRA guide nodes from ComfyUI-LTXVideo; Lightricks recommends running them at native resolution in stage 1 only.

Lightricks says most LTX-2.3 style LoRAs load on LTX-2.5, but the official IC-LoRAs have separate 2.5 versions, and the 2.3 IC-LoRAs do not work on 2.5. Always pick the file with ltx-2.5 in its name. Community LoRAs for 2.5 are still few, so test any LoRA at strength 0.6 to 1.0 with a fixed seed. Check the licence of each LoRA before commercial use.

Prompting LTX-2.5 (Video and Audio)

• Text to video: one flowing paragraph with shot type, scene, action, characters and camera movement. Example: “Wide shot of a knight in battered armour charging across a rain-soaked battlefield at dusk, the camera tracking alongside at low height, mud spraying from the horse’s hooves, thunder rolling in the distance and steel clashing.”
• Image to video: describe only what changes: motion, camera and sound. Do not re-describe what is already in the image. “Use the provided start image as the first frame” helps anchor it.
• First and last frame (FLF2V): describe the transition between the two images, and use the same aspect ratio for both.
• Audio: name the sounds and dialogue you want. If you write nothing about sound, you get generic ambience.
• Prompt style matters. Lightricks notes that prompt following depends heavily on style: concrete, visual language works better than adjectives like “epic” or “amazing”.

Need a strong start image? Make one with Qwen-Image 2.1 Turbo on 16 GB, then animate it with the I2V template.

Troubleshooting

• 401 error when downloading: you have not accepted the licence on the Lightricks repo while signed in, or your Hugging Face token is missing.
• Red “missing node” boxes in the template: ComfyUI is too old. Update it (or switch to the nightly) and restart.
• Out of memory on 16 GB: use Q3_K_M, lower stage 1 to 768×512, cut frames to 81, raise --reserve-vram, or skip the upscale stage. Check that system RAM is not full as well.
• Video but no sound: the audio VAE is missing or not connected.
• Error about projection weights or wrong encoder shape: you loaded a plain Gemma 4 12B. Use a file with “with-proj-ltx-2.5” or the DeepNeuralNerd LTX_2.5 file.
• Washed-out or burned frames: CFG is above 1 on the distilled model, or a distill LoRA is stacked on the distilled checkpoint.
• “gemma4” not supported in CLIPLoader (GGUF): the GGUF encoder needs the patch from the elix3r page. Use the INT8 safetensors encoder instead.
• More fixes: ComfyUI errors and fixes.

FAQ

Can LTX-2.5 run on 16 GB VRAM?

Yes, with the GGUF transformer (Q4_K_M 15.69 GB or Q3_K_M 12.92 GB) and enough system RAM, ideally 64 GB. The official INT8 files need about 32 GB of VRAM.

Can LTX-2.5 run on 12 GB?

Only with heavy compromises: a Q2_K or Q3_K_S GGUF, the GGUF Gemma encoder, 768×512 and short clips. Wan 2.2 at GGUF Q3 is the easier choice on 12 GB; see our Wan 2.2 16 GB guide.

How many steps does LTX-2.5 need?

The distilled model uses a fixed 8 steps at CFG 1.

What text encoder does LTX-2.5 use?

Gemma 4 12B with LTX’s projection layers in the same file. LTX-2 and 2.3 used Gemma 3 12B with a separate connector file.

Is LTX-2.5 uncensored?

ComfyUI has no content filter, but the Gemma 4 encoder can weaken some prompts. The Heretic encoder from DeepNeuralNerd removes most of that while keeping the projections.

Do LTX-2.3 LoRAs work on LTX-2.5?

Many style LoRAs do, according to Lightricks, but the official IC-LoRAs have separate 2.5 versions that you must use instead.

Can I use LTX-2.5 commercially?

Yes if your company makes under $10M a year, under the LTX-2.x Community License. Above that you need a paid licence.

Final Thoughts

LTX-2.5 is the most complete open video model of 2026: video and audio in one pass, an 8-step distilled checkpoint and a growing set of official LoRAs. On a 16 GB card the setup that works is the Abiray GGUF at Q4_K_M or Q3_K_M, an INT8 or Heretic encoder that ComfyUI parks in RAM, 960×544 at 121 frames, then the x2 upscaler. Add Slow-Motion Control or Cinemagraph for effects, keep CFG at 1, and describe sound as carefully as the picture.

Related guides:

• Wan 2.2 in ComfyUI on 8 to 16 GB
• Uncensored Local AI Video Generation in ComfyUI
• Best Local AI Video Models in 2026
• Speed Up ComfyUI on RTX 30 and 40 Cards
• ComfyUI Model Folder Cheat Sheet

Sources

• Lightricks/LTX-2.5 model card and file list
• ComfyUI docs: LTX-2.5 templates
• Abiray/LTX-2.5-Distilled-GGUF and realrebelai/LTX-2.5_GGUFs
• elix3r Gemma 4 LTX-2.5 encoder GGUF
• DeepNeuralNerd LTX-2.5 Heretic encoder
• Lightricks/ComfyUI-LTXVideo and city96/ComfyUI-GGUF
• LTX 2.5 VRAM requirements
• Lightricks LoRA model cards on Hugging Face (Slow-Motion Control, Cinemagraph, IC-LoRAs), and TheDivergentAI Turbo Distill LoRA
• Hugging Face model search for “LTX-2.5”, sorted by downloads (checked November 2, 2026)