StormForgeAI-Krea2-Turbo-LoRA-Merger

KO | EN

gitlite — search

#api #javascript #tensorflow #ai #markdown #deep-learning #chatbot #github #swift #agents

StormForgeAI-Krea2-Turbo-LoRA-Merger

(★ 9)

StormForge Krea2 Turbo LoRA Merger Suite

파일 탐색기

ZIP 다운로드

- check_aio.bat
- check_comfy_quant.bat
- check_comfy_quant.py
- check_fp8_scale.bat
- check_fp8_scale.py
- check_keys_full.py
- check_lokr_scale.bat
- check_lokr_scale.py
- check_lokr_shapes.bat
- check_lokr_shapes.py
- check_metadata.bat
- check_metadata.py
- check_specific.py
- diagnose.bat
- readme.md
- klm1.png
- klm2.png
- klm3.png
- readme.md
.gitignore
app.py
install.bat
LICENSE
lora_merge.py
README.md
requirements.txt
run.bat

# CDN으로 사용하기

jsDelivr

jsDelivr는 공개 GitHub 리포지토리를 별도 설정 없이 CDN으로 즉시 서빙합니다. 버전과 파일을 고르면 웹페이지에 바로 붙일 수 있는 링크와 예시 코드가 만들어집니다.

버전

파일

링크

예시

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

> ✅ **Confirmed working**, including split checkpoints, all-in-one
> checkpoints (diffusion model + VAE + text encoder bundled), standard LoRA
> adapters, and LyCORIS LoKr adapters (Kronecker-product format). All fixed
> and verified against real files and real renders (Fp8). If you hit issues on a
> different checkpoint/LoRA combination, please open an issue with a
> `check_keys_full.py` dump of both files.

A standalone, offline Gradio app that permanently merges one or more LoRA
files into a base diffusion model checkpoint (built for Krea 2 Turbo, but
works with any safetensors checkpoint using compatible LoRA naming).

No internet access is required to run it — everything happens locally on
your machine.

## Requirements

- Windows, with Python 3.10 or 3.11 installed and on PATH
  (https://www.python.org/downloads/ — check "Add python.exe to PATH")
- Your base checkpoint and LoRA file(s) already downloaded locally as
  `.safetensors` files

## Install

Double-click `install.bat` (or run it from a terminal). It will:

1. Create a local virtual environment in `.venv`
2. Install Gradio, PyTorch (CPU version), and safetensors

If you have an NVIDIA GPU and want faster merging, after `install.bat`
finishes, run:

```
.venv\Scripts\activate
pip install torch --index-url https://download.pytorch.org/whl/cu121
```

(swap `cu121` for the CUDA build matching your GPU driver — see
https://pytorch.org/get-started/locally/)

## Run

Double-click `run.bat`. This opens the app in your browser at
`http://127.0.0.1:7860`.

## Using the app

1. **Load files** — upload your base model `.safetensors` file, and one or
   more LoRA `.safetensors` files.
2. **LoRA weights** — a weight slider appears for each LoRA you loaded.
   `1.0` applies the LoRA at its trained strength; lower values apply it
   more subtly, negative values invert its effect.
3. **Settings**:
   - **Compute precision** — precision used for the merge math itself.
     `fp32` is safest/most accurate; `fp16`/`bf16` are faster and use less
     RAM but can lose a little precision.
   - **Output precision** — precision of the saved file. `keep original`
     preserves whatever precision the base checkpoint was in.
   - **Device** — `cpu` works everywhere; `cuda` is faster if you have a
     working GPU build of PyTorch installed.
   - **Output folder / filename** — where the merged `.safetensors` file
     is written.
4. **Check compatibility (dry run)** — before merging, run this to see how
   many LoRA layers actually match your base model's weight names. LoRAs
   trained for a different architecture, or exported in an unexpected
   format, will show up here as "unmatched" rather than silently failing.
5. **Merge & Save** — performs the actual merge and writes the output file.
   The log shows how many layers were merged, skipped, or mismatched.

## How it works

LoRA weight files store, for a subset of layers, a low-rank pair of
matrices (`down`/`up`, sometimes named `lora_A`/`lora_B`) plus an optional
`alpha` scaling value. Merging bakes each LoRA in by computing, for every
matched layer:

```
new_weight = base_weight + (up @ down) * (alpha / rank) * your_weight_slider
```

This is applied directly to the base checkpoint's tensors and the result
is saved as a new, standalone `.safetensors` file — no LoRA loader needed
at inference time.

The app recognizes both common LoRA naming conventions (Kohya/ComfyUI-style
`lora_down`/`lora_up`, and Diffusers/PEFT-style `lora_A`/`lora_B`), and
falls back to fuzzy key matching when prefixes differ between the LoRA and
base file. Any layer it can't confidently match is reported, not guessed.

## Notes / limitations

- This performs a **linear-layer LoRA merge**. It assumes standard
  `(out_features, in_features)` weight/LoRA shapes, which covers Krea 2's
  transformer architecture (and most modern DiT/UNet models). It is not
  designed for convolutional LoRAs.
- Merging loads the full base model into memory (RAM, or VRAM if using
  `cuda`). Make sure you have enough — check your checkpoint's file size
  as a rough guide.
- Always run "Check compatibility" first on a new LoRA/base combination
  before trusting the merged output.

---

## Roadmap

This project is under active development.

Planned additions include more advanced merge algorithms, additional checkpoint
utilities, model inspection tools, and workflow features aimed at making model
engineering accessible without requiring ComfyUI.

StormForgeAI-Krea2-Turbo-LoRA-Merger

오너의 다른 레포

파일 탐색기

# CDN으로 사용하기

Was this content helpful?

Recent Feedback

StormForgeAI-Krea2-Turbo-LoRA-Merger

오너의 다른 레포

파일 탐색기

# CDN으로 사용하기

Was this content helpful?

Recent Feedback

Download README