Text-to-image generation has come a long way. Fire Red Image Edit (released by Fire Red Team under Apache2.0 license) steps in with a different mindset. Instead of just making a bigger model, it focuses on optimizing everything around the model that are data, training, architecture, and evaluation.
What started as blurry textures and abstract shapes is now producing near photorealistic images with impressive detail and alignment to prompts. But there is a catch. As the quality improved, the barrier to entry shot up. Today, the ecosystem feels split into two extremes. On one side, you have got powerful restricted models like Nano Banana Pro or Seedream4.0. You do not know how they work, you cannot reproduce them, and you definitely cannot tweak them.
 |
| Fire red edit model architecture |
On the other side, the open source world is trying to keep up, mostly by scaling means bigger models, more parameters, more compute. We are talking 20B, 30B + parameter models. That's not exactly accessible for most researchers or developers.
FireRed-Image-Edit is not just trying to improve one piece of the puzzle . Its built as a full end to end system that rethinks the entire pipeline.
It starts with data, and not in a casual way. The team begins with a massive pool of around 1.6 billion samples, combining both text-to-image data and image editing pairs. But instead of blindly using all of it, the researchers aggressively clean, filter, and balance the dataset. This results in stronger semantic coverage and much better alignment between instructions and outputs.
Installation
1. Make sure you have ComfyUI installed. Older user need to update ComfyUI from the Manager itself to make all nodes working properly.
2. Download Fire Red Edit models from the list below. There are multiple model variants. Choose the one that suits your resources.
(a) Fire red Edit 1.0 (Bf16/FP8)
(b) Fire red Edit 1.0 (BF16) by ComfyUI
(c) Fire red Edit 1.0 (GGUF) by model scope
(e) Fire Red Edit 1.1 (BF16/FP8)
Save it inside ComfyUI/models/diffusion_models folder.
3. The workflow supports same text encoders, vae models that we used in the basic Qwen Image Edit 2511 workflow. If donot have then, just follow the workflow and download the text encoders and Vae from hugging face repositories.
4. Restart the ComfyUI and refresh it to take effect.
Workflow
1. Workflow is the same as of Qwen Image Edit 2511. You can also download the same workflow (Qwen_Image_Edit_2511.json) from our Hugging face repository.
2. Drag and Drop into ComfyUI. Load Fire Edit model into Load diffusion model loader. GGUF models need Unet loader node. Replace Load diffusion model loader node with this one.
3. Load other models (text encoders, vae etc ) into their respective node. Max 3 input images required.
We inputted this image into Load image loader node.
Added prompt into prompt box. Prompt- Remove the cap from Image1 and replace the outfit with red embroidered Salwar.
Here is the output that has been edited.
-Consistent editing){kind=link}