Qwen-Image-Edit (GGUF) Image Edit — Modly Extension

Instruction-based image editing with Alibaba's Qwen-Image-Edit, loaded as a 4-bit GGUF transformer so it runs on consumer GPUs. Give it one image and a text instruction and it returns edited image(s) — strong at novel viewpoints (rotate an object to show its back/side), style changes, object add/remove, and background swaps.

Apache 2.0 and ungated — no HuggingFace login, no token, no license click. The download just works.

---

VRAM reality (read this first)

Qwen-Image-Edit is a 20B model with a separate ~7B text encoder. The transformer is loaded as a low-bit GGUF, the text encoder is 4-bit, and model CPU offload swaps them on/off the GPU. The Transformer Quant is your main VRAM knob — model offload puts the whole transformer on the GPU during denoising, so it has to fit:

• ~12GB: Q3_K_M (~9GB) — the default. Q2_K (~7GB) for more headroom.
• ~10GB: Q2_K (~7GB).
• 16GB: Q4_K_M (~12GB). 24GB+: Q5_K_M / Q8_0, or Max Speed (no offload).
• 8GB: only Q2_K, and only at lower resolution — expect it to be tight.
• 6GB: not really feasible here — even the smallest quant (~7GB) is larger than the card, and the offload puts the transformer on the GPU as one piece. For genuine 6–8GB use you want the Nunchaku build (per-block offload down to ~3GB) — see Notes.

System RAM matters as much as VRAM — offloaded weights live there. 32GB+ is strongly recommended; with less you may hit out-of-memory or heavy swapping.

---

Installation

1. Open Modly → Extensions → Install from GitHub and paste this repo URL.
2. Wait for setup (installs PyTorch, diffusers, the gguf loader, and bitsandbytes into an isolated venv).
3. Click Download on the Edit Image node. It pulls two ungated repos: the 4-bit GGUF transformer (calcuis/qwen-image-edit-gguf) and the decoder bundle (callgg/image-edit-decoder, which holds the text encoder + VAE).

If the model fails to load with a from_single_file error, your diffusers is too old for Qwen GGUF loading — update it in the extension venv (pip install -U diffusers) and retry. This is bleeding-edge; the first run on a given machine sometimes needs that bump.

---

Usage (Workflows tab)

1. Drag an Image node and point it at the image you want to edit.
2. Drag an Edit Image node and type your instruction into its Edit Instruction field (e.g. rotate the object 180 degrees to show the back, three-quarter side view, change the background to a sunset beach).
3. Connect the Image node into the Edit Image node's image input.
4. Connect the Edit Image output into your Preview Image node.
5. Hit Run.

Number of Outputs = 1 emits a single image (same as the text-to-image node). Set 2–4 for multiple variations in one run (each uses a different seed); the node then emits a list of images.

Prompt tips

Qwen-Image-Edit follows explicit instructions well and preserves the subject's identity. Describe the change directly, make one change at a time, and for viewpoints be specific ("rotate 90° left", "view from above").

---

Parameters

Changing the Transformer Quant downloads that GGUF file the next time you run (the ~15GB decoder bundle is shared and isn't re-downloaded).

Memory Mode

Both modes 4-bit the text encoder via bitsandbytes (installed at setup). If it's missing, it falls back to the bf16 encoder, which needs a much larger GPU. Note: GGUF transformers are not compatible with per-layer (sequential) CPU offload, so that mode was removed — model offload is the working path.

---

Notes

• The default transformer is Q3_K_M (~9GB), chosen to fit a 12GB GPU. Use the Transformer Quant dropdown to size up (Q4/Q5/Q8) or down (Q2_K). Quants come from the GGUF repo.
• If you previously downloaded qwen-image-edit-iq4_nl.gguf, it's ~11.6GB and too big for 12GB — you can delete that file from the model folder to reclaim space.
• Want real speed and genuine 6–8GB support? Nunchaku ships 4-bit Qwen-Image-Edit with optimized CUDA kernels and per-block offload down to ~3GB. It's an extra, version-matched install, so it's left out of this build — ask for a Nunchaku variant if you want it.

---

License

Extension code is MIT — see LICENSE. The Qwen-Image-Edit weights it downloads are Apache 2.0 (Alibaba's Qwen team), which permits commercial use. This is an independent community extension, not affiliated with Alibaba, Hugging Face, or Modly; each user is responsible for complying with the model license and for whatever they generate. Provided "as is", without warranty.