Omini Qwen Image Edit Model Patch

The Flux side of Omini Kontext is the polished one. This node is the other side: the author's attempt to port the same second-image trick to Qwen-Image-Edit, the Apache-2.0 instruction editor that took the open editing crown from Kontext. It patches a Qwen-Image-Edit MODEL the same way OminiKontextModelPatch patches Flux - swapping in a forward pass that accepts an extra reference latent from OminiKontextConditioning and slots it into the image tokens with shifted position ids.

Read the README's own framing before you get excited, because it's refreshingly honest: this is an experimental model, not trained on a large dataset, and it sometimes inserts duplicate characters. This is the least-supported node in the pack, and it shows.

What you need

• A Qwen-Image-Edit checkpoint loaded as a MODEL - on a modern ComfyUI that's a standard Qwen-Image-Edit load (GGUF or fp8 to keep the 20B model sane on consumer VRAM).
• OminiKontextConditioning feeding it a VAE-encoded reference latent with the delta values that match whatever Omini-style Qwen LoRA you're using.
• The author's trained spatial-character-insertion model, which the README says lives on HuggingFace - and here's a genuine snag: the README's link is missing. It says "link below" and then there is no link. You may need to hunt for it in the repo's issues or the author's HF profile.

How it works

The patch follows the same mechanism as the Flux one: MODEL in, patched MODEL out, with the reference-aware forward and the extra_conds plumbing that carries the reference latents through to the transformer. Mechanically it's a clean port - the Omini Kontext conditioning node works with it unchanged, because the conditioning format is shared across the pack. The gap is training, not plumbing: the Qwen model hasn't been trained anywhere near as hard on the reference-insertion task as the Flux LoRAs have, which is exactly where the duplicate-character artifacts come from.

Should you use it?

Honestly? Only if you're curious or already invested in this pack. Flux Kontext's license made NSFW derivatives a violation and pushed the ecosystem to Qwen-Image-Edit - that's a big part of why the author was porting it here at all. But Qwen-Image-Edit already has a huge, actively-maintained LoRA ecosystem of its own (pose LoRAs, angle LoRAs, the works), none of which needs an Omini-style patch. This node is a research artifact in a research repo, last touched in 2025, with a missing model link and a documented artifact. If you just want "insert my character into this scene" on Qwen-Image-Edit, a standard consistency LoRA will get you further with less pain.

Install

If you do want to poke at it: ComfyUI Manager → Omini Kontext, or:

cd ComfyUI/custom_nodes
git clone https://github.com/Saquib764/omini-kontext

Restart, load Qwen-Image-Edit, patch it, and accept that duplicate characters are a feature of this particular release until the author trains a bigger model.