GenGlasses — CivitAI Model Reverse Engineer

Predicts the SD checkpoint, model family, and tags from a generated image.
Backbone: fine-tuned ViT-5.
RunPod handler: cjyu81/genglassrunpod

Repo contents

File	Description
model.safetensors	Model weights (HF-standard format)
config.json	Architecture config + training metadata
preprocessor_config.json	Image preprocessing parameters
label_maps.json	Index ↔ label maps for all three heads
training_log.json	Full run history + seen image IDs
best.pth	Legacy PyTorch checkpoint (backward compat)

Label Space

• Models : 50
• Families: 8 — Flux, Illustrious, NoobAI, Other, Pony, SD 1.5, SDXL, ZImageTurbo
• Tags : 108

Training History

Run	Date	Skip Scrape	Fresh Train	New Images	Best Val Loss
1	2026-04-04	True	True	831	2.3535

RunPod Endpoint

Important: Use RunPod Serverless → New Endpoint → Custom (Docker image), NOT "Custom Deployment" (HF model link).

{"input": {"image": "<base64-jpeg>"}}

Local usage

from safetensors.torch import load_file
from huggingface_hub import hf_hub_download
import json, torch

state = load_file(hf_hub_download('Charlie81/genglasses', 'model.safetensors'))
cfg   = json.load(open(hf_hub_download('Charlie81/genglasses', 'config.json')))
maps  = json.load(open(hf_hub_download('Charlie81/genglasses', 'label_maps.json')))
# rebuild CivitClassifier, load_state_dict(state), then predict