Deploy LTX-2.3
VideoLTX-2.3 is the fastest open-source video model with native audio generation. 22B parameters, generates synchronized video+audio in a single pass. Up to 4K resolution and 50 FPS. 18x faster than Wan 2.2.
Deploy LTX-2.3 in minutes
Starting at $0.72/hr on dedicated GPU
Available Variants (2)
| Model | GPU | VRAM | Price | Action |
|---|---|---|---|---|
LTX-2.3 22B Distilled Distilled (Recommended) | A100 80GB PCIe | 80 GB | $1.85/hr | Deploy |
LTX-2.3 22B Distilled FP8 Distilled FP8 (24GB GPU) | RTX A6000 | 48 GB | $0.72/hr | Deploy |
Prices include the service fee. Charges follow actual running time.
Requirements
ModelPilot assigns 48–80GB cloud GPUs across the listed variants. Actual local VRAM requirements vary with model variant, precision, quantization, resolution, and workflow settings.
On ModelPilot, deploy on a dedicated cloud GPU (up to 80GB VRAM) starting at $0.72/hr with no setup required.
Compare LTX-2.3
Source-backed GPU, VRAM, and cost comparisons for nearby deployment choices.
Use Cases
- ✓Fast video prototyping with audio
- ✓Social media video content
- ✓Music video generation
- ✓Product video ads
Related Models
Known Issues & Fixes
Real failures from running LTX-2.3 and similar pipelines, with the mechanism behind each one.
Output is blurry, soft, low-detail or washed-out despite a good prompt
Cause: Running a step-distilled checkpoint at classic 'quality' settings (25-30 steps, CFG 4-5). The shipped/default LTX-2.3 graphs are set up for few-step distilled inference (~8 steps, CFG 1.0, euler, explicit ManualSigmas curve), and high steps or a non-distilled schedule degrade it. Split the symptom: too many steps / non-distilled schedule -> soft/washed-out/blurry; too-high CFG on a CFG-distilled model -> burnt/oversaturated/haloed.
Fix: Use the model's few-step regime: ~8 steps, CFG 1.0, euler, leave the distilled sigma schedule in place — do NOT swap in a 25-step scheduler. (Corrected: 'shift 2.5 / LTXVScheduler' is legacy LTX-Video 0.9.x, not LTX-2.3.)
Over-processed / double-degraded output when a speed LoRA is stacked on an already-distilled checkpoint
Cause: A distill/acceleration LoRA and a pre-distilled checkpoint are alternative acceleration paths — the rule is to disable the distill LoRA whenever the checkpoint is already pre-distilled. Stacking two conflicting few-step schedules degrades output.
Fix: Apply AT MOST ONE acceleration path. Note: base Wan DOES need the lightx2v LoRA for few-step (the correct single path); the double-distill hazard only arises with a pre-merged/FusionX-style Wan checkpoint that already bakes lightx2v/CausVid in.
Model OOMs on a 4090/A6000 the file size 'should' fit; user unsure which GPU to pick
Cause: Total VRAM = diffusion weights + LLM text encoder (Gemma-3 12B, 12-24GB) + VAE + upscalers held CONCURRENTLY. Checkpoint file size alone is not the ceiling. LTX-2.3 22B bf16 ~46GB (needs 80GB), fp8 ~29GB (fits 48GB).
Fix: Size the GPU for the SUM of concurrently-loaded components; route bf16 to 80GB (A100/H100), fp8 to 48GB (A6000).
Over-caution that fp8 degrades quality, so users force bf16 and double VRAM/download for no gain
Cause: fp8_e4m3fn_scaled checkpoints carry per-block scale factors that preserve dynamic range, matching bf16 quality for many modern diffusion/video models at ~half VRAM. Plain (unscaled) fp8_e4m3fn CAN visibly degrade — that's why the scaled repacks exist.
Fix: Prefer the 'scaled' fp8 variant; don't reflexively pick bf16. Do NOT generalize the claim to unscaled fp8.
'Value not in list: gemma_path ... not in []', or the pod hard-exits on a missing tokenizer.model even though you supplied a CLIP/T5
Cause: Newer models replaced CLIP/T5 with a full LLM encoder. LTX-2 19B's LTXVGemmaCLIPModelLoader wants a DIRECTORY encoder: weights come from the non-gated Comfy-Org/ltx-2 repo, but config.json + tokenizer.model come from the GATED google/gemma-3-12b-it repo — without HF token + accepted license the sidecar 403s and the loader sees an empty list.
Fix: Supply HF token + accept the Gemma license so the gated config/tokenizer sidecar downloads into the encoder directory. Note LTX-2.3's default graph instead uses a single-file FP4 encoder (gemma_3_12B_it_fp4_mixed, non-gated, no tokenizer sidecar).
Encoder file is present on disk but the node still errors, or a workflow references a filename you 'basically' have
Cause: Loaders are picky three orthogonal ways: (1) PRECISION — Kijai LoadWanVideoT5TextEncoder wants its own bf16 umt5 build and rejects native ComfyUI's scaled fp8; native CLIPLoader(type=wan) accepts fp8. (2) PACKAGING — LTX LTXVGemmaCLIPModelLoader wants a directory encoder; LTXAVTextEncoderLoader wants a single-file fp4-mixed .safetensors. (3) EXACT NAME — t5xxl_fp8_e4m3fn.safetensors (4.89GB) and t5xxl_fp8_e4m3fn_scaled.safetensors (5.16GB) are distinct, not interchangeable.
Fix: Match the encoder file to the node's precision + packaging + exact filename; fuzzy substitution fails hard.
Hitting a specific error message? Browse ComfyUI error fixes.
Frequently Asked Questions
How much GPU memory is allocated for LTX-2.3?
The listed ModelPilot variants use 48–80GB cloud GPUs. Local memory needs vary with the variant, precision, quantization, and workflow settings.
How much does it cost to run LTX-2.3?
Starting at $0.72/hr on a dedicated GPU. Charges are calculated from actual running time, with auto-stop when credits run out.
How long does LTX-2.3 take to deploy?
Video models typically take 15–30 minutes for first deployment (model download + setup). Subsequent starts are faster.
Can I run LTX-2.3 on my local GPU?
It depends on the selected variant, precision, quantization, and workflow settings. Compare the variants below with your available VRAM; the table shows ModelPilot's cloud GPU allocation, not a universal local minimum.
Ready to deploy LTX-2.3?
Pick your GPU and have it running in minutes. No infrastructure setup required.