Skip to main content
Back

ChatR Labs Drops Omni LoRA to Handle Image Generation and Editing Together

ChatR Labs released a rank-128 LoRA that handles both text-to-image and reference-based image editing on Qwen-Image-2.1 using one adapter.

  • ChatR Labs released Omni LoRA v0.1, a single adapter for Qwen-Image-2.1 generation and editing.
  • Rank-128 BF16 adapter, 671 MB, 224 modules across 32 transformer blocks.
  • Trained to step 32,000 at 1024x1024 with aspect-ratio bucketing enabled.
  • Runs via Diffusers QwenImage21Pipeline; recommended 25 steps, true CFG 4.0, LoRA strength 1.0.
  • Playground benchmarks: 33s text-to-image, 38s single-reference edit on dual RTX 4090s.
  • Licensed under Qwen Research License, non-commercial only; GitHub repo hosts inference code.

ChatR Labs has released Omni LoRA v0.1, an experimental community adapter for Qwen-Image 2.1. The same loaded adapter handles text-to-image generation and natural-language edits, reducing weight swaps for applications that expose both workflows.

One adapter stays loaded

A LoRA is a compact set of additional weights that changes a base model’s behavior while reusing its core components. Most Qwen-Image LoRAs specialize in one task or visual style. Omni v0.1 covers generation and editing with a single rank-128 adapter.

The 671 MB download contains one BF16 safetensors file. The transformer, text encoder, processor, and variational autoencoder come from the Qwen base model. ChatR Labs has released no merged checkpoint, FP8 build, or GGUF variant.

Qwen 2.1 changes the architecture

Qwen-Image 2.1 is Alibaba’s newer 7-billion-parameter image model. Its architecture differs from earlier Qwen-Image releases, making the base checkpoint and adapter version strict compatibility requirements.

Omni v0.1 compatibility
Model or weight Compatible Reason
Qwen/Qwen-Image-2.1 Yes Training base for Omni v0.1
Earlier Qwen-Image, Qwen-Image-2512, and Qwen-Image-Edit-2511 bases No Different model architectures