mainbrains/sdxl-vfx-cinematic-lora

SDXL VFX Cinematic LoRA

A LoRA fine-tune of Stable Diffusion XL trained on high-resolution VFX reference frames and cinematic stills.

Overview

This LoRA was trained to generate images with professional VFX aesthetics — volumetric lighting, practical effects, lens artifacts, and cinematic color grading. Sourced from my own VFX reference library (~2,400 curated frames from personal compositing work and film breakdowns).

Parameter	Value
Base model	SDXL 1.0
LoRA rank	32
Training images	2,400
Resolution	1024x1024
Steps	3,000
Learning rate	1e-4
Optimizer	AdamW8bit
Trained on	RTX 4090, ~6 hours

Trigger Words

• cinematic vfx — general cinematic VFX look
• volumetric — enhanced volumetric lighting and fog
• practical fx — practical effects aesthetic (pyro, dust, debris)
• lens artifact — anamorphic flares, bokeh, chromatic aberration

Usage

ComfyUI

Load as a standard SDXL LoRA node. Recommended weight: 0.7–0.9.

Diffusers

from diffusers import StableDiffusionXLPipeline
import torch

pipe = StableDiffusionXLPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0",
    torch_dtype=torch.float16
).to("cuda")

pipe.load_lora_weights("mainbrains/sdxl-vfx-cinematic-lora")

image = pipe(
    "cinematic vfx explosion, volumetric lighting, film grain, anamorphic lens flare",
    num_inference_steps=30,
    guidance_scale=7.5
).images[0]

Recommended Settings

Setting	Value
Sampler	DPM++ 2M Karras
Steps	25–35
CFG	6.5–8.0
LoRA weight	0.7–0.9
Resolution	1024x1024 or 1216x832

Limitations

• Best results at 1024x1024 or standard SDXL resolutions
• May oversaturate at LoRA weights above 1.0
• Trained on English captions only

License

CreativeML OpenRAIL-M — same as SDXL base.