Fixing OOM Errors in ComfyUI with LTX-2.3 and NVIDIA RTX 5060 Ti

The NVIDIA RTX 5060 Ti with 16GB of GDDR7 VRAM sits at an interesting price point for AI video generation. It’s powerful enough to run modern models, but right at the edge of what’s needed for LTX-2.3 video generation. If you’re hitting out-of-memory (OOM) errors, you’re not alone—and you’re not out of luck.

This guide covers the exact settings, quantization options, and workflow adjustments that make LTX-2.3 work reliably on your RTX 5060 Ti 16GB.

Understanding the Memory Challenge

RTX 5060 Ti 16GB Specifications

Specification	Value
VRAM	16GB GDDR7
CUDA Cores	4,608
Memory Bandwidth	672 GB/s
Architecture	Blackwell (GB206-300)
Process	5nm
TDP	180W

The 16GB buffer puts this card in a challenging spot for LTX-2.3. The full-precision model demands around 46GB of VRAM—nearly three times what’s available. But Blackwell’s fifth-generation Tensor Cores unlock precision formats that change the equation dramatically.

LTX-2.3 Memory Requirements by Model Variant

Model Variant	VRAM Required	RTX 5060 Ti Compatible?
FP32 (full)	~46GB	❌ No
FP16	~23GB	❌ No
FP8	~23GB	⚠️ Borderline
GGUF Q4	~18GB	✅ Yes (tight)
GGUF Q4_0 Distilled	~12.3GB	✅ Yes (recommended)

The distilled GGUF Q4_0 variant is your best bet. It fits comfortably within 16GB with room for the rest of your workflow.

Memory allocation breakdown for LTX-2.3 on RTX 5060 Ti 16GB

Essential ComfyUI Launch Arguments

Start ComfyUI with these arguments to optimize memory for LTX-2.3:

python main.py --lowvram --reserve-vram 5 --fp8-e5m2-unet

What Each Flag Does

--lowvram: Splits the UNET model into smaller chunks and offloads parts to system RAM. This is essential for 16GB cards running LTX-2.3. You’ll see slightly longer generation times, but it prevents crashes.

--reserve-vram 5: Reserves 5GB of VRAM for the operating system and other processes. Without this, ComfyUI might try to use every last megabyte, making your system unstable.

--fp8-e5m2-unet: Uses FP8 precision for the UNET model. Blackwell GPUs handle FP8 natively with excellent quality retention.

Alternative Launch Configurations

If you’re still hitting OOM:

# More aggressive memory management
python main.py --lowvram --reserve-vram 6 --use-split-cross-attention

# Last resort (very slow but stable)
python main.py --novram --reserve-vram 2

Model Selection: Choosing the Right Variant

Recommended: GGUF Q4_0 Distilled

Download the distilled Q4_0 variant of LTX-2.3:

ltx-video-2.3-distilled-Q4_0.gguf

This model runs at approximately 12.3GB VRAM with SageAttention enabled, leaving headroom for:

• VAE decode
• Tokenizer and text encoders
• System overhead

Why Not FP8?

FP8 models technically fit on 16GB cards, but the margins are thin. Any additional nodes in your workflow— ControlNet adapters, upscalers, or secondary models—can push you over the limit. GGUF Q4 gives you more breathing room.

SageAttention Integration

SageAttention reduces memory consumption by 30-35% and speeds up generation. Install it:

pip install triton sageattention

In your ComfyUI workflow, add a “Patch Sage Attention” node and set the backend to:

sageattn_qk_int8_pv_fp16_cuda

Don’t use the --use-sage-attention CLI flag—it can cause compatibility issues with some models. The node-based approach is more reliable.

Resolution Limits and Safe Operating Ranges

Starting resolution matters more than output resolution. LTX-2.3 generates video from latents, and the latent size directly impacts VRAM consumption.

Safe Resolution Guidelines

Safety Level	Starting Resolution	Notes
Conservative	512 × 384	Always works
Moderate	768 × 512	May need VAE offload
Aggressive	1024 × 640	Requires all optimizations

Workflow Strategy

1. Generate at lower resolution first (512 × 384)
2. Upscale afterward using RTX Video Super Resolution or another upscaler
3. Keep frame count low (25-50 frames to start)

This two-pass approach is faster and more reliable than trying to generate at high resolution directly.

VAE Optimization Techniques

Use Tiled VAE Decode

OOM most commonly occurs during the VAE decode step—the final stage where latents become video frames. Replace your standard VAE Decode node with “VAE Decode (Tiled)” and set:

tile_size: 256
overlap: 64

Lower tile sizes use less VRAM but take longer. Start at 256 and increase if you have VRAM headroom.

VAE Precision Settings

For best quality on RTX 50-series:

--fp32-vae    # Best quality, more VRAM

For memory-constrained situations:

--fp16-vae    # Saves ~2GB VRAM

Blackwell handles FP16 well, so the quality difference is minimal.

Managing Long Video Projects

LTX-2.3 on 16GB VRAM has limits. If you need longer videos, generate in segments:

Chunked Generation Method

1. Generate 2-3 seconds (50-75 frames) at a time
2. Save the last frame as an image
3. Use the last frame as the first-frame input for the next segment
4. Stitch segments together with ffmpeg:

ffmpeg -f concat -i segments.txt -c copy output.mp4

Creating Frame Continuity

In ComfyUI, use an “Image Saver” node or “Math Expression” node to pass the last generated frame as input to your next generation. This maintains visual continuity across chunks.

Common OOM Scenarios and Fixes

”CUDA Out of Memory” During Model Load

Cause: Model too large for available VRAM after system reservation.

Fix: Switch to GGUF Q4_0 distilled model, or use:

--lowvram --reserve-vram 6

“Out of Memory” During VAE Decode

Cause: Latent tensor size exceeds VAE decode capacity.

Fix: Use VAE Decode (Tiled) with tile_size: 256.

Random Stalls or System Freezes

Cause: ComfyUI consuming all VRAM, starving the display.

Fix: Increase --reserve-vram to 6-8 GB. Yes, you’re “wasting” VRAM, but your system remains stable.

”Kernel launch failed” Error

Cause: PyTorch trying to allocate more memory than available.

Fix: This usually means you don’t have enough system RAM for offloading. Ensure you have at least 32GB system RAM, and close other applications.

Batch Size vs Batch Count

Understanding this distinction prevents many OOM errors:

Batch Size (in Empty Latent Image node): Number of samples processed simultaneously. Higher = more VRAM.

Batch Count (generation iterations): Number of sequential generations. Does NOT increase VRAM.

Recommendation for RTX 5060 Ti

• Batch Size: Always 1
• Batch Count: As high as you want

Processing multiple generations sequentially is safer than trying to parallelize them.

RTX 50-Series Specific Advantages

Blackwell architecture brings real advantages for LTX-2.3:

FP8 Native Support

Fifth-generation Tensor Cores handle FP8 with no performance penalty compared to FP16. On older architectures (Ampere, Ada), FP8 requires conversion overhead. On Blackwell, it’s native.

NVFP4 Format

For future-proofing, NVFP4 offers:

• 60% VRAM reduction over FP16
• 2.5× faster inference

Currently, NVFP4 model variants are limited, but this will change as the ecosystem matures.

Weight Streaming

ComfyUI on RTX 50-series can utilize system RAM more efficiently through weight streaming. Enable it in ComfyUI settings under “Memory Management” → “Enable weight streaming.”

Complete Working Configuration

Here’s a known-good configuration for RTX 5060 Ti 16GB:

Launch Command

python main.py --lowvram --reserve-vram 5 --fp8-e5m2-unet --fp32-vae

Model

ltx-video-2.3-distilled-Q4_0.gguf

Resolution

Width: 512
Height: 384
Frame Count: 49 (2 seconds at 24fps)

SageAttention Node

Backend: sageattn_qk_int8_pv_fp16_cuda

VAE Decode (Tiled)

tile_size: 256
overlap: 64

Empty Latent Image

batch_size: 1

This configuration should generate 2-second video clips reliably. From there, you can experiment with higher resolutions or longer frame counts—just add one change at a time to isolate what breaks.

When to Upgrade

If you consistently need:

• Resolution above 1024 × 640: Consider RTX 5070 Ti or higher (20GB+ VRAM)
• Frame counts above 100: Multi-chunk generation works, but patience wears thin
• Multiple ControlNet adapters: Each adapter adds VRAM overhead

The RTX 5060 Ti 16GB handles LTX-2.3 admirably for its price point. With quantized models and thoughtful workflow design, you can generate high-quality video content without the crashes.

Troubleshooting Checklist

Before asking for help, verify:

• Using GGUF Q4_0 distilled model variant
• ComfyUI launched with --lowvram
• VRAM reserved for system (--reserve-vram 5)
• Batch size set to 1
• Resolution starts at 512 × 384 or lower
• VAE Decode uses Tiled mode
• SageAttention installed and configured via node
• System RAM ≥ 32GB
• All GPU applications closed (browser, games, video players)

Resources

• NVIDIA RTX AI Garage Blog — Official LTX integration guide
• ComfyUI LTX-Video Plugin — Required custom nodes
• ComfyUI Wiki GPU Guide — Hardware recommendations
• SageAttention GitHub — Performance optimizations

---

Last updated: March 2026. Settings may evolve with ComfyUI and LTX-2.3 updates.