by InstantX
Open source · 6k downloads · 321 likes
FLUX.1 dev IP Adapter is an AI model designed to integrate visual references into image generation by combining the capabilities of the FLUX.1-dev model with an image adaptation system (IP Adapter). It enables users to guide image creation by leveraging visual examples, providing greater control over style, colors, or elements to be reproduced. Primarily used for artistic generation, design, or visual content creation, it stands out for its ability to treat images as text, simplifying their integration into creative workflows. However, its performance varies depending on the use case, with limitations in fine style transfer or character consistency, sometimes requiring multiple attempts to achieve satisfactory results.
This repository contains a IP-Adapter for FLUX.1-dev model released by researchers from InstantX Team, where image work just like text, so it may not be responsive or interfere with other text, but we do hope you enjoy this model, have fun and share your creative works with us on Twitter.
This is a regular IP-Adapter, where the new layers are added into 38 single and 19 double blocks. We use google/siglip-so400m-patch14-384 to encode image for its superior performance, and adopt a simple MLPProjModel of 2 linear layers to project. The image token number is set to 128. The currently released model is trained on the 10M open source dataset with a batch size of 128 and 80K training steps.
We adopt Shakker-Labs/FLUX.1-dev-LoRA-collections as a character LoRA and use its default prompt.
The code has not been integrated into diffusers yet, please use our local files at this moment.
import os
from PIL import Image
import torch
import torch.nn as nn
from pipeline_flux_ipa import FluxPipeline
from transformer_flux import FluxTransformer2DModel
from attention_processor import IPAFluxAttnProcessor2_0
from transformers import AutoProcessor, SiglipVisionModel
from infer_flux_ipa_siglip import resize_img, MLPProjModel, IPAdapter
image_encoder_path = "google/siglip-so400m-patch14-384"
ipadapter_path = "./ip-adapter.bin"
transformer = FluxTransformer2DModel.from_pretrained(
"black-forest-labs/FLUX.1-dev", subfolder="transformer", torch_dtype=torch.bfloat16
)
pipe = FluxPipeline.from_pretrained(
"black-forest-labs/FLUX.1-dev", transformer=transformer, torch_dtype=torch.bfloat16
)
ip_model = IPAdapter(pipe, image_encoder_path, ipadapter_path, device="cuda", num_tokens=128)
image_dir = "./assets/images/2.jpg"
image_name = image_dir.split("/")[-1]
image = Image.open(image_dir).convert("RGB")
image = resize_img(image)
prompt = "a young girl"
images = ip_model.generate(
pil_image=image,
prompt=prompt,
scale=0.7,
width=960, height=1280,
seed=42
)
images[0].save(f"results/{image_name}")
Please refer to ComfyUI-IPAdapter-Flux.
You can also enjoy this model at Shakker AI.
This model supports image reference, but is not for fine-grained style transfer or character consistency, which means that there exists a trade-off between content leakage and style transfer. We don't find similar properties in FLUX.1-dev (DiT-based) as in InstantStyle (UNet-based). It may take several attempts to get satisfied results. Furthermore, current released model may suffer from limited diversity, thus cannot cover some styles or concepts,
The model is released under flux-1-dev-non-commercial-license. All copyright reserved.
This project is sponsored by HuggingFace, fal.ai and Shakker Labs.
If you find this project useful in your research, please cite us via
@misc{flux-ipa,
author = {InstantX Team},
title = {InstantX FLUX.1-dev IP-Adapter Page},
year = {2024},
}