Update README.md

README.md

@@ -9,4 +9,112 @@ library_name: diffusers
 tags:
 - text-generation-inference
 - image-edit
----
+---
+
+# **FireRed-Image-Edit-1.0-fp8**
+
+> **FireRed-Image-Edit-1.0-fp8** is an FP8-compressed transformer variant built on top of **FireRedTeam/FireRed-Image-Edit-1.0**.
+> This release provides **Transformers-only compressed weights** and **Diffusers-compatible transformer weights**, enabling reduced memory usage and improved throughput while preserving the high-fidelity instruction-based image editing capabilities of the original model.
+
+> [!important]
+> This release compresses **only the diffusion transformer module** using **BF16 · FP8 (F8_E4M3)** precision. The VAE and other components remain unchanged from the base model. FP8 (8-bit floating point) weight and activation quantization using hardware acceleration on GPUs –
+FP8 W8A8: [https://docs.vllm.ai/en/stable/features/quantization/fp8/](https://docs.vllm.ai/en/stable/features/quantization/fp8/)
+Quantization recipe: [https://github.com/vllm-project/llm-compressor/tree/main/examples/quantization_w8a8_fp8](https://github.com/vllm-project/llm-compressor/tree/main/examples/quantization_w8a8_fp8)
+
+## Diffusers Usage
+
+```python
+import torch
+from diffusers.models import QwenImageTransformer2DModel
+from diffusers import QwenImageEditPlusPipeline
+from diffusers.utils import load_image
+
+transformer = QwenImageTransformer2DModel.from_pretrained(
+    "prithivMLmods/FireRed-Image-Edit-1.0-fp8",
+    subfolder="transformer",
+    torch_dtype=torch.bfloat16
+)
+
+pipeline = QwenImageEditPlusPipeline.from_pretrained(
+    "FireRedTeam/FireRed-Image-Edit-1.0",
+    transformer=transformer,
+    torch_dtype=torch.bfloat16
+)
+
+pipeline.to("cuda")
+
+image1 = load_image("grumpycat.png")
+prompt = "turn the cat into an orange cat"
+
+inputs = {
+    "image": [image1],
+    "prompt": prompt,
+    "generator": torch.manual_seed(42),
+    "true_cfg_scale": 1.0,
+    "negative_prompt": " ",
+    "num_inference_steps": 40,
+    "guidance_scale": 1.0,
+    "num_images_per_prompt": 1,
+}
+
+output = pipeline(**inputs)
+output_image = output.images[0]
+output_image.save("output_image_edit_plus.png")
+```
+
+## About the Base Model
+
+**FireRed-Image-Edit-1.0** from FireRedTeam is a state-of-the-art open-source diffusion transformer designed for instruction-based image editing.
+
+It achieves top-tier performance through:
+
+* A **1.6B-sample dataset**, refined to **100M+ high-quality text-to-image and editing pairs**
+* Cleaning, stratification, auto-labeling
+* Dual-stage filtering for optimal semantic coverage and instruction alignment
+
+### Multi-Stage Training Pipeline
+
+1. Pre-training
+2. Supervised fine-tuning
+3. Reinforcement learning
+
+### Key Innovations
+
+* **Multi-Condition Aware Bucket Sampler** for efficient variable-resolution batching
+* **Stochastic Instruction Alignment** with dynamic prompt re-indexing
+* **Asymmetric Gradient Optimization** for stable DPO
+* **DiffusionNFT** with layout-aware OCR rewards for precise text editing
+* **Differentiable Consistency Loss** for identity preservation
+
+## Native Capabilities
+
+* Photo restoration
+* Multi-image editing such as virtual try-on
+* Style transfer with text fidelity
+* Complex instruction adherence
+* Layout-aware text editing
+* Identity-preserving edits
+* Professional photorealistic refinements
+
+  * Skin texture realism
+  * Multi-outfit changes in single passes
+
+It achieves strong results across:
+
+* REDEdit-Bench with 15 editing categories
+* ImgEdit
+* GEdit
+
+The model supports native editing from text-to-image foundations rather than patch-based methods, enabling coherent, high-quality outputs suitable for professional workflows and ComfyUI integration.
+
+## What FP8 Adds
+
+The **FireRed-Image-Edit-1.0-fp8** variant introduces:
+
+* **BF16 · FP8 (F8_E4M3) Transformer Compression**
+* Reduced VRAM usage
+* Improved throughput
+* Faster inference on Hopper and compatible GPUs
+* Production-friendly deployment without modifying the original pipeline structure
+
+> Only the transformer weights are compressed, ensuring seamless compatibility with existing Diffusers pipelines.