LTX 2.5 ComfyUI Setup Guide
LTX 2.5 lives in a gatedHuggingFace repo. Sign in and accept the license first, then pick a transformer and Gemma 4 text encoder for your VRAM. If you're on 16GB, the official files won't fit — you need the community GGUF path.
1. Install ComfyUI
Clone the ComfyUI repository and install dependencies. Requires Python 3.10+ and a CUDA-capable GPU.
git clone https://github.com/comfyanonymous/ComfyUI cd ComfyUI pip install -r requirements.txt
2. Accept the LTX 2.5 license (gated repo)
Every official LTX 2.5 file lives in Lightricks/LTX-2.5 (license: ltx-2-community-license-agreement), which is gated. Sign in to HuggingFace, open the repo, click "Agree and Access," and only then will file downloads succeed — an anonymous download attempt returns 401. The separate IC-LoRA pixel upscaler repo is gated the same way.
3. Download a transformer
Pick one. Official BF16/INT8/NVFP4 files require sign-in from step 2; community GGUF quants (last four rows) are not gated but need the ComfyUI-GGUF custom node to load.
| File | Size | Target folder | Gated |
|---|---|---|---|
ltx-2.5-22b-dev-transformer-bf16.safetensors | 42.02 GB | ComfyUI/models/checkpoints/ | 🔒 Yes |
ltx-2.5-22b-dev-transformer-comfy-int8-convrot.safetensors | 21.50 GB | ComfyUI/models/checkpoints/ | 🔒 Yes |
ltx-2.5-22b-distilled-transformer-bf16.safetensors | 42.02 GB | ComfyUI/models/checkpoints/ | 🔒 Yes |
ltx-2.5-22b-distilled-transformer-comfy-int8-convrot.safetensors | 21.50 GB | ComfyUI/models/checkpoints/ | 🔒 Yes |
ltx-2.5-22b-distilled-transformer-nvfp4.safetensors | 18.72 GB | ComfyUI/models/checkpoints/ | 🔒 Yes |
LTX-2.5-Distilled-Q3_K_S.gguf | 12.65 GB | ComfyUI/models/checkpoints/ | No |
LTX-2.5-Distilled-Q4_K_M.gguf | 15.69 GB | ComfyUI/models/checkpoints/ | No |
LTX-2.5-Distilled-Q6_K.gguf | 18.62 GB | ComfyUI/models/checkpoints/ | No |
LTX-2.5-Distilled-Q8_0.gguf | 23.60 GB | ComfyUI/models/checkpoints/ | No |
4. Download the required VAEs
Video VAE decodes latents to frames; video VAE conv is a second variant shipped alongside it; audio VAE is for audio-video generation.
| File | Size | Target folder | Gated |
|---|---|---|---|
ltx-2.5-video-vae-bf16.safetensors | 1.47 GB | ComfyUI/models/vae/ | 🔒 Yes |
ltx-2.5-video-vae-conv-bf16.safetensors | 1.45 GB | ComfyUI/models/vae/ | 🔒 Yes |
ltx-2.5-audio-vae-bf16.safetensors | 0.36 GB | ComfyUI/models/vae/ | 🔒 Yes |
5. Download the Gemma 4 12B text encoder
Every LTX 2.5 workflow needs one of these. The projection layer is bundled in, unlike LTX 2.3's separate text-projection file. On 16GB, only the GGUF Q5_K_M fits alongside a GGUF transformer.
| File | Size | Target folder | Gated |
|---|---|---|---|
gemma4-12b-with-proj-ltx-2.5-bf16.safetensors | 26.26 GB | ComfyUI/models/text_encoders/ | 🔒 Yes |
gemma4-12b-with-proj-ltx-2.5-comfy-int8-convrot.safetensors | 15.37 GB | ComfyUI/models/text_encoders/ | 🔒 Yes |
gemma4-12b-with-proj-ltx-2.5-Q5_K_M.gguf | 9.51 GB | ComfyUI/models/text_encoders/ | 🔒 Yes |
6. Optional components
Distillation LoRA, latent upscalers for two-stage pipelines, the duration-head patch, and the IC-LoRA pixel upscaler.
| File | Size | Target folder | Gated |
|---|---|---|---|
ltx-2.5-22b-distilled-lora-450-bf16.safetensors | 8.90 GB | ComfyUI/models/loras/ | 🔒 Yes |
ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors | 1.00 GB | ComfyUI/models/latent_upscale_models/ | 🔒 Yes |
ltx-2.5-latent-temporal-upscaler-x2-bf16-1.0.safetensors | 0.26 GB | ComfyUI/models/latent_upscale_models/ | 🔒 Yes |
ltx-2.5-duration-head-bf16.safetensors | 3.84 MB | ComfyUI/models/model_patches/ | 🔒 Yes |
ltx-2.5-22b-ic-lora-pixel-spatial-upscaler-x2-1.0.safetensors | 0.327 GB | ComfyUI/models/loras/ | 🔒 Yes |
7. What's different from LTX 2.3
- Text encoder is Gemma 4 12B, not Gemma 3 12B — and the projection layer is bundled into the same file. LTX 2.3 needed a separate ltx-2.3_text_projection_bf16.safetensors; LTX 2.5 does not.
- The official ComfyUI quant format is int8-convrot (filenames contain "comfy"), not FP8 scaled. There is no official FP8 build and no Kijai fork for LTX 2.5 — Lightricks/LTX-2.5 is the only official source.
- The smallest official text encoder is 15.37GB (Gemma 4 INT8) versus 9.5GB for LTX 2.3's Gemma 3 FP4. Combined with the smallest official transformer (18.72GB NVFP4 or 21.50GB INT8), an all-official pipeline needs well over 24GB — 16GB cards must use community GGUF files instead.
- A new duration-head model patch ships with LTX 2.5, with no LTX 2.3 equivalent — it belongs in ComfyUI/models/model_patches/.
- A second video VAE variant (video-vae-conv) ships alongside the standard video VAE, with no LTX 2.3 equivalent.
- The official distillation LoRA is rank 450, up from rank 384 in LTX 2.3.
- The Lightricks/LTX-2.5 repo is gated — you must sign in to HuggingFace and click "Agree and Access" before any file will download. LTX 2.3's repo is not gated.
8. Loading a workflow
We haven't independently verified specific ComfyUI node graphs or example workflow JSON filenames for LTX 2.5 yet, so we won't list any here and risk sending you after a file that doesn't exist. Check the file listing on the gated Lightricks/LTX-2.5 repo and the ComfyUI-LTXVideo node repo for the current example graphs, and confirm your node version explicitly supports LTX 2.5 before wiring anything up. We'll add a verified workflow section here once we can confirm exact filenames.