LTX23_video_vae_bf16.safetensors
LTX 2.3 Video VAE (Kijai)
Full-precision BF16 video VAE by Kijai — the heavier, higher-fidelity alternative to the tiny taeltx2_3 decoder. Use it when a workflow asks for it or when you need the cleanest decode on long or HDR clips. Place in models/vae/.
Released 2026-05-27 · Source: Kijai/LTX2.3_comfy (HuggingFace) — Added in Kijai's late-May component batch alongside the audio VAE and text projection.
Download LTX23_video_vae_bf16.safetensors
Direct HuggingFace download. ~1 GB · Free.
No 2GB GPU? Try LTX23_video_vae_bf16.safetensors online — free generation included
Skip the ~1 GB download and ComfyUI setup. Generate a 6-second video using this exact model in your browser, ~30 seconds.
Technical details
LTX23_video_vae_bf16.safetensors is the full BF16 video VAE for LTX 2.3, as opposed to taeltx2_3.safetensors which is the distilled Tiny AutoEncoder (TAE). Both decode latents to frames and encode frames to latents (for I2V); the difference is fidelity vs cost. The full VAE keeps more dynamic range and detail, which matters most on long sequences, high-contrast / HDR content, and two-stage upscale pipelines where decode artifacts compound.
It is roughly 1 GB BF16 and lives in ComfyUI/models/vae/ — the same folder as taeltx2_3 and the audio VAE. They coexist; you select which one a given workflow uses in its VAELoader node.
In audio-video pipelines this is the video-side VAE that pairs with LTX23_audio_vae_bf16.safetensors. For ordinary silent T2V/I2V it is optional — taeltx2_3 is the default and is what almost every published workflow references.
When to choose LTX23_video_vae_bf16.safetensors
Pick the full video VAE in three cases: (1) a workflow JSON explicitly references LTX23_video_vae_bf16.safetensors and the node turns red with the tiny VAE; (2) you are running audio-video / talking-head workflows, where it pairs with the audio VAE; (3) you are getting visible banding, color drift, or mushy detail from taeltx2_3 on long or HDR clips and want a cleaner decode.
For everything else, use taeltx2_3.safetensors. It is smaller, faster, and is the filename hardcoded in the vast majority of community and official workflows. Downloading the full VAE 'just in case' is unnecessary.
The two are interchangeable at the node level — you can keep both in models/vae/ and switch per workflow without re-downloading.
Will this run on my GPU?
Minimum: 2GB VRAM.
Recommendation: Most setups only need taeltx2_3.safetensors. Download this full BF16 VAE only if a workflow references it by name, or if you see banding/color drift with the tiny VAE on long clips.
How to use LTX23_video_vae_bf16.safetensors
- Download the file from HuggingFace.
- Place it in ComfyUI/models/vae/ inside your ComfyUI directory.
- Restart ComfyUI (or refresh the model list from the menu).
- Load a compatible workflow — see below.
Don't want to run this locally? Try LTX23_video_vae_bf16.safetensors online with a free generation — no GPU, no install, ~30 seconds per clip.
ComfyUI says it can't find LTX23_video_vae_bf16.safetensors?
Some published workflow JSONs reference this file under a custom subdirectory. If ComfyUI shows a "cannot find model" error and your workflow references one of these path-prefixed variants:
- vae/LTX23_video_vae_bf16.safetensors
- ltx23_video_vae_bf16_kj.safetensors
- ltx23\LTX23_video_vae_bf16.safetensors
The prefix before the slash or backslash is a subdirectory the workflow author used. The actual file is the same LTX23_video_vae_bf16.safetensors — you have two fixes:
- Create the matching subdirectory inside ComfyUI/models/vae/ and place the file there. Example: if the workflow references vae/LTX23_video_vae_bf16.safetensors, create the corresponding subfolder under ComfyUI/models/vae/ and put LTX23_video_vae_bf16.safetensors inside it.
- Or open the workflow JSON in a text editor and replace the prefixed string with just LTX23_video_vae_bf16.safetensors. ComfyUI then resolves it directly from ComfyUI/models/vae/.
On Windows the separator is \, on macOS/Linux it is / — they refer to the same nested folder regardless of platform.
Common issues
VAELoader: 'LTX23_video_vae_bf16.safetensors not in list'▼
File placed in ComfyUI/models/ root or in checkpoints/ instead of vae/. Fix: Move it to ComfyUI/models/vae/LTX23_video_vae_bf16.safetensors and click refresh on the VAELoader node so ComfyUI re-scans.
Workflow references 'ltx23_video_vae_bf16_kj.safetensors' but my file has a different name▼
Some workflow authors rename the file with a '_kj' suffix to mark it as Kijai's. The file is identical; only the referenced string differs. Fix: Either rename your local copy to match the workflow, or edit the VAELoader node and reselect LTX23_video_vae_bf16.safetensors from the dropdown.
Decoded video is solid green / black / corrupt▼
Partial download — HF's xet protocol can leave a truncated file that still presents a valid .safetensors header. Fix: Delete and redownload via `huggingface-cli download` or `aria2c` with retry; verify the size matches HuggingFace's reported ~1 GB.
ComfyUI doesn't see the file after I downloaded it▼
Make sure the file is in ComfyUI/models/vae/ (not a subfolder). Restart ComfyUI fully — the menu refresh sometimes misses new files. Filename must match exactly: LTX23_video_vae_bf16.safetensors.
CUDA out of memory error when loading the model▼
LTX23_video_vae_bf16.safetensors needs ~2GB VRAM minimum. If you're hitting OOM: • Enable Sequential Offloading in ComfyUI settings • Lower the resolution (768×512 instead of 1280×704) — both dimensions must be divisible by 32 • Reduce frame count (65 frames instead of 161) — must be 8n+1 • Use a smaller variant — see Related models below.
How do I apply this LoRA in ComfyUI?▼
Load it in a 'LoraLoader' node and connect it after your model loader. Pair this LoRA with the dev base model (not the distilled one) for the right behavior. LoRA strength 1.0 is the trained value — start there.
Get notified when LTX 2.3 Video VAE (Kijai) updates
Occasional updates on what's new in LTX 2.3 — new FP8 quants, LoRAs, IC-LoRA releases — with our hands-on verdict on whether they're worth re-downloading. No fixed cadence.
No spam. Sent occasionally when there's real news. Unsubscribe in one click.
Related models
ltx-2.3-22b-distilled-1.1_lora-dynamic_fro09_avg_rank_111_bf16.safetensors
ltx-2.3-22b-distilled-lora-1.1_fro90_ceil72_condsafe.safetensors
LTX-2.3-OmniNFT-RL-Lora_bf16.safetensors
ltx-2.3-22b-distilled-lora-384-1.1.safetensors
ltx-2.3-22b-distilled-lora-384.safetensors
ltx-2.3-22b-distilled-lora-dynamic_fro09_avg_rank_105_bf16.safetensors