tensorart/stable-diffusion-3.5-large-TurboX

TensorArt-TurboX-SD3.5Large

<div align="center"> <img src="ckptad.webp" alt="ComfyUItemppsqad00205_" width="400"/> </div>

Model Description

TensorArt-TurboX-SD3.5Large is a highly optimized variant of Stable Diffusion 3.5 Large, achieving 6x faster generation speed with minimal quality loss. It surpasses the official Stable Diffusion 3.5 Large Turbo in image detail, diversity, richness, and realism.

The model is available in both LoRA and checkpoint (ckpt) formats(chosen one of them, don't use them at same time), ensuring compatibility with most community models. This allows seamless integration into various workflows, accelerating open-source AI development.

Key Features

• —🚀 6x Speed Boost: Generates images significantly faster than the original SD3.5 Large.
• —🎨 Superior Quality: Outperforms stabilityai/stable-diffusion-3.5-large-turbo in detail, diversity, and realism.
• —🔄 Versatile Compatibility: Works with both realistic and anime-style models.
• —⚡ Optimized for Efficiency: Suitable for both high-end and mid-range GPUs.

Recommended Settings

For optimal results, we recommend using:

• —Sampler: Euler
• —Scheduler: Simple
• —Sampling Steps: 8
• —CFG Scale: 1.0 - 1.5 (Note: CFG ≠ 1 will double generation time)
• —Shift: 5 (Most important！！！)
• —LoRA Strength: 1.0 (if using LoRA version)

Example Comparisons

TensorArt-TurboX-SD3.5Large vs. Official Models

Here are some sample outputs comparing TensorArt-TurboX-SD3.5Large (8 steps, CFG=1, Simple scheduler) with official stable-diffusion-3.5-large-turbo when using the same time. The contrast workflow are here: constrast_normal_TurboX, contrast_TurboX_turbo:

Example Output

Using ckpt：

    import torch
    from diffusers import StableDiffusion3Pipeline

    repo = "tensorart/stable-diffusion-3.5-large-TurboX"


    pipe = StableDiffusion3Pipeline.from_pretrained(repo, torch_dtype=torch.float16)
                                                    
    pipe = pipe.to("cuda")


    generator = torch.Generator(device="cuda").manual_seed(1)

    image = pipe(
    "A beautiful bald girl with silver and white futuristic metal face jewelry, her full body made of intricately carved liquid glass in the style of Tadashi, the complexity master of cyberpunk, in the style of James Jean and Peter Mohrbacher. This concept design is trending on Artstation, with sharp focus, studio-quality photography, and highly detailed, intricate details.",
    num_inference_steps=8,
    guidance_scale=1.5,
    height=1024,
    width=768,
    generator=generator
    ).images[0]

    image.save("./test2-3.webp")

Using lora：

   import torch
   from diffusers import StableDiffusion3Pipeline, FlowMatchEulerDiscreteScheduler
   import numpy as np
   from safetensors.torch import load_file
   from huggingface_hub import hf_hub_download

   repo = "tensorart/stable-diffusion-3.5-large-TurboX"
   ckpt = "Tensorart_TurboX_sd3.5L_8steps.safetensors"

   pipe = StableDiffusion3Pipeline.from_pretrained("stabilityai/stable-diffusion-3.5-large", torch_dtype=torch.float16,)
                                                   
   pipe = pipe.to("cuda")
   pipe.load_lora_weights(hf_hub_download(repo, ckpt))
   pipe.fuse_lora()
   
   pipe.scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained(repo, subfolder="scheduler", shift=5)

   generator = torch.Generator(device="cuda").manual_seed(1)
   image = pipe(
      "A beautiful bald girl with silver and white futuristic metal face jewelry, her full body made of intricately carved liquid glass in the style of Tadashi, the complexity master of cyberpunk, in the style of James Jean and Peter Mohrbacher. This concept design is trending on Artstation, with sharp focus, studio-quality photography, and highly detailed, intricate details.",
      num_inference_steps=8,
      guidance_scale=1.5,
      height=1024,
      width=768,
      generator=generator
   ).images[0]
   image.save("./test1.webp")

Community Model Integration

TensorArt-TurboX-SD3.5Large seamlessly integrates with various community models:

• —✅ Photorealistic Models: Enhances realism while maintaining the original style.
• —✅ Anime-Style Models: Preserves artistic style while boosting generation speed.
• —✅ LoRA Models: Works with style, character, and concept LoRAs.

Limitations

While highly optimized, this model may not be suitable for:

1. 1.Rendering precise English text in images.
2. 2.Generating anatomically perfect hands.
3. 3.Experiments with non-recommended samplers or schedulers (Best performance with Euler + Simple scheduler).

Model Links

• —TensorArt (LoRA Version): TensorArt-TurboX-SD3.5Large-8steps

Contact

• —Website: https://tensor.art https://tusiart.com
• —Developed by: TensorArt