SYSTMS ACTION: QWEN IMAGE EDIT 2511
Turn just about anything into a toy with the ACTION Qwen Image Edit 2511 LoRA! Simply upload an image and prompt it by writing "action the [subject]"!
Try an example
About this workflow
This ComfyUI tutorial workflow demonstrates instruction-based image editing with Qwen Image Edit 2511 and a targeted "ACTION" LoRA that turns subjects into toy-like versions. It loads the Qwen Image Edit 2511 diffusion model with UNETLoader and applies the ACTION LoRA via LoraLoaderModelOnly to bias the aesthetic toward playful, miniature, or collectible-toy looks. The QwenEditTextEncode_EditUtils node pairs a dedicated Qwen 2.5 VL text encoder with your input image and instruction (for example, "action the cat") to create edit-aware conditioning, while RandomNoise initializes the latent.
SamplerCustomAdvanced runs the diffusion process using your chosen sampler from KSamplerSelect and the schedule from BasicScheduler, with CFGGuider controlling classifier-free guidance strength.
The input image is prepared for the model using ResizeImagesByLongerEdge and ImageCrop+ to standardize dimensions, with GetImageSize and PrimitiveInt passing size info to the sampler. After sampling, VAELoader and VAEDecode convert latents back to an RGB image, and SaveImage writes the result. Because this is an instruction-following edit model, you don’t need an init-latent path—your original photo is consumed by the Qwen encoder as a visual reference while the LoRA steers the result toward a consistent toy style.
The included Note node documents the prompt pattern ("action the [subject]") so you can quickly produce consistent outputs.
Questions & answers
Download the Qwen Image Edit 2511 diffusion model (UNET) and the Qwen 2.5 VL text encoder from the links provided. Place the diffusion model in ComfyUI/models/diffusion_models and the text encoder in ComfyUI/models/text_encoders. Load a compatible VAE in VAELoader. The ACTION LoRA is loaded via LoraLoaderModelOnly.
Increase or decrease the ACTION LoRA strength in LoraLoaderModelOnly and tune CFG in CFGGuider. Higher LoRA strength and higher CFG generally intensify the toy effect; lowering them preserves more of the original look. You can also adjust steps for detail (more) vs. speed (fewer).
No. This workflow uses instruction-based editing: the input image and text instruction are encoded together by QwenEditTextEncode_EditUtils, and the diffusion starts from noise. Control intensity via LoRA strength, guidance (CFGGuider), sampler choice, and steps.
Start with a balanced sampler like DPM++ 2M Karras or Euler and 20–30 steps. The ResizeImagesByLongerEdge and ImageCrop+ nodes set a consistent canvas—use a size appropriate for your GPU. Higher resolutions need more VRAM and time; if you see OOM errors, lower the long edge size or steps.