LTX-2.3 (22B) on 12 GB of VRAM
Nothing fits entirely, but Q2_K (8.3 GB) overflows by only about 1.1 GB. ComfyUI keeps that part in system RAM automatically: slower than a full fit, but usable.
01What to download
| Part | File | Folder | Size | |
|---|---|---|---|---|
| Model | ltx-2.3-22b-dev-Q2_K.gguf Q2_K · unsloth/LTX-2.3-GGUF | models/unet | 8.3 GB | Download → |
| Text encoder | gemma_3_12B_it_fp8_scaled.safetensors Gemma 3 12B FP8 (scaled) · Comfy-Org/ltx-2 | models/text_encoders | 13.2 GB | Download → |
| VAE | ltx-2.3-22b-dev_video_vae.safetensors LTX-2.3 video VAE · unsloth/LTX-2.3-GGUF | models/vae | 1.5 GB | Download → |
| Also needed | ltx-2.3-22b-dev_audio_vae.safetensors LTX-2.3 audio VAE · unsloth/LTX-2.3-GGUF | models/vae | 0.4 GB | Download → |
| Also needed | ltx-2.3-22b-dev_embeddings_connectors.safetensors LTX-2.3 text embeddings connectors (projection) · unsloth/LTX-2.3-GGUF | models/text_encoders | 2.3 GB | Download → |
| Total download · keep about the same free on disk | 25.6 GB | |||
Alternatives: Gemma 3 12B FP4 mixed (9.4 GB); Gemma 3 12B BF16 (24.4 GB). The 16-bit text encoder is a little more faithful; the smaller one is chosen here because it loads faster and needs less RAM. FP4 files are listed only as alternatives unless the GPU is an RTX 50 (Blackwell). Only for some workflows: LTX-2.3 spatial upscaler x2 1.1 (1.0 GB, official two-stage template). Sizes read from Hugging Face (2026-09-25). “Download” links start the file directly; the file name links are the same files the official ComfyUI workflows use.
System RAM: 32 GB or more. ComfyUI keeps the model file, the text encoder and the VAE in system RAM and moves them to the GPU as needed. With this set of files that is about 25.6 GB, plus roughly 6 GB for Windows, ComfyUI and a browser: 31.6 GB in total. With less RAM it still runs, but Windows starts swapping to disk and loading gets very slow. calculated
02Every file, on 12 GB
| File | Size | Needed | On this card | Quality | Download |
|---|---|---|---|---|---|
| 16-bit GGUF · unsloth | 42.0 GB | 46.8 GB | 34.8 GB too big | the original weights | Hugging Face → |
| Q8_0 GGUF · unsloth | 22.8 GB | 27.6 GB | 15.6 GB too big | practically identical to the original | Hugging Face → |
| Q6_K GGUF · unsloth | 17.8 GB | 22.6 GB | 10.6 GB too big | very close to the original | Hugging Face → |
| Q5_K_M GGUF · unsloth | 16.1 GB | 20.9 GB | 8.9 GB too big | close; small differences in fine detail | Hugging Face → |
| Q4_K_M GGUF · unsloth | 14.3 GB | 19.1 GB | 7.1 GB too big | good; some loss in fine detail and text | Hugging Face → |
| Q3_K_M GGUF · unsloth | 10.8 GB | 15.6 GB | 3.6 GB too big | noticeable loss of detail | Hugging Face → |
| Q2_K ← GGUF · unsloth | 8.3 GB | 13.1 GB | spills 1.1 GB | heavy loss; a last resort | Hugging Face → |
Sizes from the Hugging Face file listing, read 2026-09-25. “Needed” = file + 4 GB working memory for this model + 0.8 GB kept free for the system.
03The text encoder
Gemma 3 12B: 24.4 GB as 16-bit, 13.2 GB as FP8, 9.4 GB as FP4 mixed. ComfyUI encodes the prompt first and can push the encoder out of VRAM before sampling, so it does not have to fit together with the model. Use the smallest encoder file; it still fits on its own.