cahlen/acid-ansi-lora

ACID ANSI LoRA

A LoRA fine-tune of Stable Diffusion XL 1.0 that generates images in the style of classic ANSI art from the BBS era (1990s). Trained on rendered ANSI/RIP art from the ACiD Productions art packs.

Usage

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("cahlen/acid-ansi-lora")

prompt = "acid-ansi-style, a menacing skull wreathed in flames on a black background"
negative_prompt = "blurry, photorealistic, photo, smooth gradients, 3d render, watermark"

image = pipe(
    prompt,
    negative_prompt=negative_prompt,
    num_inference_steps=30,
    guidance_scale=7.5,
    cross_attention_kwargs={"scale": 0.7},
).images[0]

image.save("output.png")

Trigger Word

acid-ansi-style — prepend this to your prompts.

Recommended Settings

Parameter	Value
Inference steps	30
Guidance scale	7.5
LoRA scale	0.6 - 0.8 (0.7 recommended)
Negative prompt	blurry, photorealistic, photo, smooth gradients, 3d render, watermark

LoRA Scale

The LoRA scale controls style strength at inference time. Lower values (0.5-0.6) preserve more of the base model's composition while higher values (0.8-1.0) push harder into the ANSI style but may cause repetition artifacts. 0.7 is the recommended default.

Samples

All samples generated at step 3500, LoRA scale 0.7, seed 42.

Skull	Dragon

BBS Login	Cityscape

Baseline Comparison (No LoRA)

Without LoRA	With LoRA (scale 0.7)

Training Details

Parameter	Value
Base model	stabilityai/stable-diffusion-xl-base-1.0
LoRA rank	32
LoRA alpha	32
Optimizer	Prodigy (lr=1.0, constant scheduler)
Training steps	3500
Batch size	1 (gradient accumulation 4)
Precision	bf16
SNR gamma	5
Noise offset	0.05
Caption dropout	0.15
Training framework	SimpleTuner

Training Data

276 images derived from 50 curated ANSI/RIP art renders from ACiD Productions art packs (1993-1996):

• Nearest-neighbor upscaling to preserve blocky pixel edges
• Multi-crop extraction for tall/narrow ANSI art (overlapping 1440px sections)
• VGA 16-color palette quantization as data augmentation
• Instance prompt only (no per-image captions)