LTX Desktop currently hard-gates local inference at 32GB VRAM, forcing any GPU below that threshold into API-only mode. This includes the RTX 4090 (24GB VRAM), which is one of the most capable consumer GPUs on the market.
The 32GB requirement appears to stem from loading both the video generation model and the Gemma text encoder into VRAM simultaneously. However, the text encoder can run on CPU/system RAM without meaningful performance impact, as demonstrated by every other local inference implementation of LTX-2 (ComfyUI, the community LTX-2-OPTIMIZED repo, diffusers pipeline with enable_sequential_cpu_offload, etc.).
This creates a situation where the same model runs perfectly on the same hardware through other tools, but LTX Desktop refuses to even attempt local inference and silently routes users to paid API generation.
This is especially confusing given that Lightricks marketing materials (The Neuron launch coverage, ltx.io/ltx-desktop landing page) advertise local inference support down to 8GB VRAM on RTX 3070 laptops.
Requested changes:
-Implement CPU offloading for the text encoder when VRAM is below 32GB but the GPU is otherwise CUDA-capable
-At minimum, lower the VRAM threshold or make it configurable so users with capable hardware can opt in to local inference
-If the concern is stability, present it as a user choice rather than a hard gate; let users with 24GB cards try local inference and fall back to API if it fails, rather than blocking them outright
Environment: Windows 11, RTX 4090 24GB, 64GB system RAM, i9-14900K
The current behavior contradicts both the open-source ethos of the project and the marketing claims about consumer hardware support.
LTX Desktop currently hard-gates local inference at 32GB VRAM, forcing any GPU below that threshold into API-only mode. This includes the RTX 4090 (24GB VRAM), which is one of the most capable consumer GPUs on the market.
The 32GB requirement appears to stem from loading both the video generation model and the Gemma text encoder into VRAM simultaneously. However, the text encoder can run on CPU/system RAM without meaningful performance impact, as demonstrated by every other local inference implementation of LTX-2 (ComfyUI, the community LTX-2-OPTIMIZED repo, diffusers pipeline with enable_sequential_cpu_offload, etc.).
This creates a situation where the same model runs perfectly on the same hardware through other tools, but LTX Desktop refuses to even attempt local inference and silently routes users to paid API generation.
This is especially confusing given that Lightricks marketing materials (The Neuron launch coverage, ltx.io/ltx-desktop landing page) advertise local inference support down to 8GB VRAM on RTX 3070 laptops.
Requested changes:
-Implement CPU offloading for the text encoder when VRAM is below 32GB but the GPU is otherwise CUDA-capable
-At minimum, lower the VRAM threshold or make it configurable so users with capable hardware can opt in to local inference
-If the concern is stability, present it as a user choice rather than a hard gate; let users with 24GB cards try local inference and fall back to API if it fails, rather than blocking them outright
Environment: Windows 11, RTX 4090 24GB, 64GB system RAM, i9-14900K
The current behavior contradicts both the open-source ethos of the project and the marketing claims about consumer hardware support.