LTX-2.5 (22B) on 20 GB of VRAM
Download Q4_K_M (15.1 GB). With the model's working memory it needs about 19.9 GB, leaving 0.1 GB spare on 20 GB. Quality: good; some loss in fine detail and text. For better quality, Q5_K_M (16.8 GB) also runs, with about 1.6 GB spilling into system RAM — a little slower, still practical.
01What to download
| Part | File | Folder | Size | |
|---|---|---|---|---|
| Model | LTX-2.5-Distilled-Q4_K_M.gguf Q4_K_M · realrebelai/LTX-2.5_GGUFs | models/unet | 15.1 GB | Download → |
| Text encoder | gemma4-12b-with-proj-ltx-2.5-comfy-int8-convrot.safetensors Gemma 4 12B + LTX projection INT8 (convrot) · Lightricks/LTX-2.5 | models/text_encoders | 15.4 GB | Download → |
| VAE | ltx-2.5-video-vae-bf16.safetensors LTX-2.5 video VAE (DiffVAE) BF16 · Lightricks/LTX-2.5 | models/vae | 1.5 GB | Download → |
| Also needed | ltx-2.5-audio-vae-bf16.safetensors LTX-2.5 audio VAE + vocoder BF16 · Lightricks/LTX-2.5 | models/vae | 0.4 GB | Download → |
| Total download · keep about the same free on disk | 32.3 GB | |||
Alternatives: Gemma 4 12B + LTX projection BF16 (26.3 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. Only for some workflows: LTX-2.5 latent spatial upscaler x2 (1.0 GB, official two-stage template); Gemma 4 E2B INT8 (prompt enhancer, optional) (5.2 GB, only if prompt enhancer enabled). 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: 48 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 32.3 GB, plus roughly 6 GB for Windows, ComfyUI and a browser: 38.3 GB in total. With less RAM it still runs, but Windows starts swapping to disk and loading gets very slow. calculated
02Every file, on 20 GB
| File | Size | Needed | On this card | Quality | Download |
|---|---|---|---|---|---|
| 16-bit SAFETENSORS · Lightricks | 42.0 GB | 46.8 GB | 26.8 GB too big | the original weights | Hugging Face → |
| Q8_0 GGUF · realrebelai | 23.6 GB | 28.4 GB | 8.4 GB too big | practically identical to the original | Hugging Face → |
| Q6_K GGUF · realrebelai | 18.7 GB | 23.5 GB | 3.5 GB too big | very close to the original | Hugging Face → |
| Q5_K_M GGUF · realrebelai | 16.8 GB | 21.6 GB | spills 1.6 GB | close; small differences in fine detail | Hugging Face → |
| Q4_K_M ← GGUF · realrebelai | 15.1 GB | 19.9 GB | fits · 0.1 GB spare | good; some loss in fine detail and text | Hugging Face → |
| Q3_K_M GGUF · realrebelai | 11.5 GB | 16.3 GB | fits · 3.7 GB spare | noticeable loss of detail | Hugging Face → |
| Q2_K GGUF · realrebelai | 8.8 GB | 13.6 GB | fits · 6.4 GB spare | 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 4 12B (custom LTX-2.5 build): 26.3 GB as 16-bit, 15.4 GB as INT8. 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. On 20 GB the INT8 encoder fits on its own, so prompt encoding stays fast.