💻 Hardware Requirements¶
The LTX 2.3 values below are tested minimums for the listed settings. The LTX 2.5 values are conservative provisioning targets based on the official model sizes and ComfyUI offloading; they are not universal benchmarked minimums. Higher resolution, longer videos, larger batches, concurrent model loading, or less offloading can require more VRAM and system RAM.
LTX-2.5¶
GPU and system RAM¶
| Configuration | Typical GPU | VRAM target | System RAM | Guidance |
|---|---|---|---|---|
| int8-convrot transformer and encoder | RTX 3090, RTX 4090, RTX 5090 | 24 GB minimum; 32 GB preferred | 64 GB minimum; 80 GB recommended | LVRAM model set; use ComfyUI offloading. |
| int8-convrot transformer with BF16 encoder | L40S, RTX A6000, RTX 6000 Ada | 48 GB | 80 GB minimum; 96 GB recommended | HVRAM model set used by an L40S with the supplied threshold. |
| int8-convrot transformer with BF16 encoder | A100/H100 80 GB, RTX PRO 6000 96 GB | 80 GB or more | 96 GB minimum; 128 GB recommended | HVRAM model set with more headroom for larger workflows. |
RunPod VRAM selection¶
The supplied LTX 2.5 profiles use VRAM_THRESHOLD=40. The startup script
selects the HVRAM model set when detected VRAM is greater than 40 GB. An
L40S therefore receives the comfyicu Dev INT8 ConvRot transformer with the
Heretic BF16 text encoder. GPUs at or below the threshold receive the same Dev
transformer with the Heretic INT8 ConvRot text encoder.
Storage¶
| Selected chain | Model storage guidance |
|---|---|
| Public Dev INT8 ConvRot with Heretic encoder | Allow at least 60 GB |
| Private dev int8-convrot, distilled LoRA, and Heretic encoder | Allow at least 65 GB |
Use at least 120 GB of persistent /workspace storage for one complete profile,
ComfyUI, custom nodes, and a modest amount of output. Increase this for long or
high-resolution videos because generated frames, previews, and cached files can
quickly consume the remaining space.
LTX 2.3¶
The example screenshots in the overview page document successful LTX 2.3 runs with these GPU classes on RunPod.
BF16¶
- Precision: bf16.
- Video settings: 1920x1088, 20 sec, 24 fps.
| GPU | Min VRAM | Min RAM |
|---|---|---|
| L40S / RTX 6000 Ada | 45 GB | 60 GB |
FP8¶
- Precision: fp8 mixed.
- Video settings: 1280x736, 20 sec, 24 fps.
| GPU | Min VRAM | Min RAM |
|---|---|---|
| RTX A5000 / RTX 4090 | 24 GB | 50 GB |