🎬 LTX-2.3 NVFP4 Quantized Models
Compress LTX-2.3 from 46GB → 13GB while maintaining most of the quality
These are NVFP4 quantized variants of LTX-2.3, optimized for users with limited VRAM who still want high-quality video generation. Along with the models, I'm sharing Comfy Bathroom — a custom node suite designed specifically for using LoRAs with these heavily quantized models.
📦 Available Models
| Model | Size | Description |
|---|---|---|
LTX-2.3-FP4 |
~21GB | Standard NVFP4 quantization |
LTX-2.3-FP4ME |
~13GB | Mixed Extreme - Blocks 0, 47 in BF16; Block 1, 46 in FP8 |
📦 Coming Soon Models
| Model | Size | Description |
|---|---|---|
LTX-2.3-FP4MEL |
~15GB | Mixed Extreme LoRA-friendly |
⚠️ Note: FP4ME variants have known LoRA compatibility issues. Use the included Comfy Bathroom nodes for best results.
🚿 Comfy Bathroom — Custom Node Suite
Comfy Bathroom is a LoRA loading system designed to work with NVFP4 overly quantized models. It provides per-block weight control to avoid artifacts when using LoRAs on aggressively quantized models.
Why "Bathroom"?
Because I made the FP4 models with Comfy Kitchen, and when something goes wrong in the kitchen... you end up in the bathroom. 🚽
🪥 Nodes Included
| Node | Icon | Description |
|---|---|---|
| Toothbrush | 🪥 | LoRA loader with built-in FP4 presets |
| Mirror (Simple) | 🪞 | Per-block on/off toggles |
| Mirror (Fancy) | 🪞 | Per-block 0.0-1.0 strength sliders |
| Shower | 🚿 | Quick preset applier |
| Towel | 🧾 | LoRA packet info display |
| Bathroom Sink | 🚰 | Stack multiple LoRAs with global strength |
🔧 How It Works
The Problem
When using LoRAs on NVFP4 quantized models, certain transformer blocks cause artifacts:
- Low blocks (0-10) → Smoky/cloudy output (noise generation interference)
- High blocks (40-47) → Double vision/ghosting (detail alignment issues)
- Middle blocks (11-39) → Usually stable for LoRA injection
The Solution
Comfy Bathroom applies intelligent weight curves to transformer blocks, ramping up LoRA strength gradually in early blocks and tapering down in late blocks.
Built-in Presets
| Preset | Description | Best For |
|---|---|---|
| FP4ME Light | Gentle ramp up/down, block 1 OFF | Style LoRAs, general use |
| FP4ME Heavy | Aggressive taper to 0 at block 46 | Heavy style modifications |
| FP4MEL Light | Offset for BF16 blocks 1, 46 | FP4MEL model + style LoRAs |
| FP4MEL Heavy | Aggressive taper with BF16 safe zones | FP4MEL + content LoRAs |
Weight Curve Visualization
FP4ME Light:
Block: 0 1 2 3 4 5 6 7 8 9 10 11 ... 39 40 41 42 43 44 45 46 47
Weight: 1.0 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 1.0 1.0 ... 1.0 0.95 0.9 0.85 0.8 0.75 0.7 0.65 0.5 1.0
BF16 FP8 ←── RAMP UP ──→ FULL ←────── TAPER DOWN ──→ FP8 BF16
📸 Example Workflow
Basic Usage
Toothbrush (LoRA + Preset) ──► Bathroom Sink ──► Model
│
global_strength
Advanced Usage
Toothbrush ──► Shower (swap preset) ──► Mirror (fine tune) ──► Bathroom Sink ──► Model
│ │
└──────────────────────────────────────────────────────────────┘
Multiple LoRAs can be stacked
💡 Usage Tips
For Style LoRAs (90% of use cases)
- Use FP4ME Light preset
- The base model already knows anatomy/structure
- LoRA only needs to affect the "style layer" (middle blocks)
For Content LoRAs (specific characters, unusual anatomy)
- Try FP4ME Heavy preset
- May need to experiment with Mirror nodes for fine-tuning
- Content LoRAs teach new concepts → need more block coverage
For Multiple LoRAs
- Connect multiple Toothbrush nodes to Bathroom Sink
- Use global_strength to scale everything together
- Individual LoRA strengths multiply with global
📋 Requirements
- ComfyUI (latest version recommended)
comfy_kitchen(for NVFP4 model loading)- BF16/FP16/FP8/FP4 capable GPU (known to work on 3xxx and above)
- 16GB+ VRAM recommended, seen people use it on 12GB
📁 Installation
- Download model files to
ComfyUI/models/checkpoints/*unless you have a working unet loader for ltx2.3 - Install Comfy Bathroom:
ComfyUI/custom_nodes/comfy_bathroom.py - Restart ComfyUI
🙏 Credits
- LTX-2.3 by Lightricks
- NVFP4 Quantization via Modified Comfy Kitchen
- Comfy Bathroom Nodes written by GLM5 (Z.ai)
📜 License
This model is subject to the LTX-2 Community License Agreement.
When your LoRA is full of artifacts, head to the Bathroom. 🚿
- Downloads last month
- 116