RealVisXL V5 Turbo (FP8 Standalone Safetensors)
This repository contains a standalone FP8 (float8_e4m3fn) quantized weights conversion of RealVisXL V5 Turbo (originally created by SG161222).
This version is specifically converted and optimized for budget, mobile, or legacy hardware setups—particularly 4GB and 6GB VRAM GPUs (such as the GTX 1650 Ti, GTX 1660, or RTX 3050 Mobile). By reducing the model's footprint, it allows the UNet execution workspace to stay entirely within physical VRAM, preventing aggressive system RAM offloading and severe pagefile disk thrashing.
⚡ Hardware Benefits
- VRAM Footprint: Reduced to ~2.4 GB, leaving adequate headroom for high-resolution activation maps.
- Inference Speed: Drastically cuts down iteration times on low-end hardware. On a legacy GTX 1650 Ti (4GB), 768x768 generations complete in sub-30 seconds once text encoders are cached.
🛠️ Recommended Inference Settings
Because this is a highly distilled Turbo model, traditional SDXL generation rules do not apply. Pushing step counts too high will over-correct tensors and degrade the output image.
| Parameter | Recommended Value |
|---|---|
| Resolution | 768x768 (Optimal for speed/VRAM balancing) |
| Inference Steps | 4 to 8 steps (6 steps is the sweet spot) |
| CFG Scale | 1.0 – 2.0 (Keep it low to avoid frying the distillation layers) |
| Sampler | dpmpp_2m or euler_ancestral |
| Scheduler | turbo or sgm_uniform |
🚀 How to Use in ComfyUI
- Download the
.safetensorsfile from this repository. - Place the file into your local ComfyUI directory:
ComfyUI/models/checkpoints/orComfyUI/models/diffusion_models/depending on your specific workflow node structure. - In your workflow, swap out your standard checkpoint loader for this FP8 file.
- Ensure your ComfyUI startup flags include
--lowvramor--fp8_e4m3fn-textencif you want to compress your pipeline even further.
⚠️ Note on Quality: While FP8 drastically speeds up generation on lower-end cards, distillation combined with 8-bit precision can cause a minor loss in high-frequency micro-textures (like fine skin pores or complex fabric weaves) compared to native BF16 non-turbo models. For high-fidelity production finishes, pair this base generation with a dedicated neural upscaler node (e.g., 4x-UltraSharp).
👥 Credits & Attribution
All credit for the underlying model tuning goes to SG161222, the original creator of the RealVisXL series. If you appreciate the composition and realism capabilities of this model, please check out their original work and support their community contributions.