Note: This repository hosts my personal management scripts/tools only. No model weights are included. No LoRA training/tuning features are provided here.
📊 waveshaker/dataset_collections
🎨 waveshaker/universal-lora-studio
models_manager_v*.py is used for centralized local management of waveshaker/loras_gallery | waveshaker/loras | waveshaker/dataset_collection: (Click to expand)Features:
waveshaker/loras_gallery/models.json, maintain card metadata (Title, Trigger, Dataset URL, Showcase URL, Update Time in Unix timestamp), and globally manage custom KV fields. Supports both full-prefix JSON (for local editing) and clean relative-path JSON (for minimal cloud uploads)..safetensors and .webp files to the waveshaker/loras Model repo. Automatically appends new model entries with an update_time timestamp and updates the minimal cloud waveshaker/loras_gallery/models.json.waveshaker/dataset_collections Dataset repo.gallery.json, and push to the waveshaker/loras/comfyui_showcase folder in the Models repo. Supports both local-only generation and incremental upload.
base_model_uploader_gui.py is a GUI tool for batch uploading base models and cloning remote repos to waveshaker/comfy_models (Click to expand)Features:
.safetensors files. It supports multi-threaded transfers, auto-resumption, and status tracking to skip completed files.comfy_dataset_prep_wf_v**.json: A ComfyUI workflow for LoRA dataset prep. (Click to expand)Integrates Krea2, Qwen Image 2.1, Flux 2 Klein 9B, and SQVLM for high-res(1-1.56MP) upscaling and auto-captioning of faces and bodies.
my LLM settings in system_prompts_user.json in ComfyUI\custom_nodes\ComfyUI_Simple_Qwen3-VL-gguf(aka SQVLM):
"📄🖼️🎬Qwen3.8-27B-Uncensored-HauhauCS-Aggressive.IQ3_XS": {
"model_path": "G:\\StableDiffusion\\MODELS\\LLM\\GGUF\\Qwen3.8-27B-Uncensored-HauhauCS-Aggressive.IQ3_XS.gguf",
"mmproj_path": "G:\\StableDiffusion\\MODELS\\LLM\\GGUF\\Qwen3.8-27B-Uncensored-HauhauCS-Aggressive.mmproj-BF16.gguf",
"max_tokens": 4096,
"max_frames": 81,
"image_min_tokens": 128,
"image_max_tokens": 1024,
"n_ctx": 102400,
"n_batch": 2048,
"n_ubatch": 512,
"enable_thinking": false,
"n_keep": 512,
"n_gpu_layers": -1,
"n_threads": 8,
"use_mmap": true,
"split_mode": 0,
"temperature": 0.8,
"top_p": 0.95,
"min_p": 0.05,
"repeat_penalty": 1.1,
"presence_penalty": 0.0,
"top_k": 40,
"chat_handler": "qwen35",
"script": "qwen3vl_run.py",
"debug": true,
"verbose": false
},
showcase_tempalte_v**.json: A ComfyUI workflow designed to generate PNG images with embedded workflow metadata. (Click to expand)The accompanying Python tool models_manager_v*.py parses these images to extract the exact Prompt, generation parameters, and LoRA stacks, automatically compiling them into a structured gallery.json for the web gallery's "ComfyUI Showcase" viewer.
Wildcards used in workflow:
filter_nsfw.py: A script for rough NSFW filtering(based on NudeNet). It provides a quick initial screen, but manual review is still necessary for accurate results.
batch_preview_generator/: A 5-steps batch pipeline for auto-generating and deploying showcase preview images. (Click to expand)It follows a structured 5-step workflow:
0_batch_template_*.json to batch generate random preview images.1_batch_regroup.py to automatically reorganize all output images into their respective subfolders.2_batch_process_meta.py to extract the meta data from the png files and convert the images into webp format.3_batch_upload_showcase.py to batch upload the processed showcase folders to Hugging Face.4_batch_modify_json.py to automatically update the models.json data by regex-matching the showcase items for each entry, leaving you ready for a final manual upload.The batch_preview_generator/ is primarily designed for adapting older models like Pony and FLUX.1 - Dev, with the main goal of rapid batch generation. For newer models, use showcase_tempalte_v**.json to generate images, and then upload them using models_manager_v*.py.
hf_space_cap_checker.py: Compares and calculates the total storage size of models defined in waveshaker/comfy_models/studio_config.json and waveshaker/universal-lora-studio/model_config.py. Filters duplicates via the HF API to ensure the total size does not exceed Hugging Face's 150 GB capacity limit.
checkpoint_splitter/: Utilities to split AIO checkpoints into VAE, TE, and UNET, including UNET int8 quantization and checkpoint re-merging. (Click to expand)check_keys.py: Inspects the keys of split safetensors files to verify and adjust split logic.checkpoint_int8_convrot.py: Uses Kijai's quant_int8_convrot.py to quantize the UNET of an AIO checkpoint to int8 convrot, without altering the Text Encoder (TE) or VAE.split_checkpoint_Illustrious.py: Extracts the VAE, TE, and UNET from Illustrious checkpoints.split_checkpoint_pony.py: Extracts the VAE, TE, and UNET from Pony checkpoints.split_checkpoint_sd15.py: Extracts the VAE, TE, and UNET from SD 1.5 checkpoints.split_checkpoint_sdxl.py: Extracts the VAE, TE, and UNET from SDXL checkpoints.