--- license: apache-2.0 language: - zh - en - fr - ko - ja - de - it - pt base_model: - black-forest-labs/FLUX.1-schnell pipeline_tag: text-to-image library_name: diffusers --- ![image](./chinese.png) ![FLUX.1 [schnell] Grid](./PEA-Diffusion.png) `MultilingualFLUX.1-adapter` is a multilingual adapter tailored for the Flux.1 series models, theoretically, it inherits ByT5 and can support over 100 languages, but with additional optimizations in Chinese. Originating from an ECCV 2024 paper titled [PEA-Diffusion](https://arxiv.org/abs/2311.17086). The open-source code is available at https://github.com/OPPO-Mente-Lab/PEA-Diffusion. # Usage We used the multilingual encoder [byt5-xxl](https://huggingface.co/google/byt5-xxl/tree/main), and the teacher model used in the adaptation process was FLUX.1-schnell. We implemented a reverse denoising process for distillation training. The adapter can be applied to any FLUX.1 series model in theory. Here we provide the following application examples. ## `MultilingualFLUX.1` The same applies to other FLUX.1 series models, just remember to adjust the `num_inference_steps` and `guidance_scale` as needed. ```python from diffusers import FluxPipeline, AutoencoderKL from diffusers.image_processor import VaeImageProcessor from transformers import T5ForConditionalGeneration,AutoTokenizer import torch import torch.nn as nn class MLP(nn.Module): def __init__(self, in_dim=4096, out_dim=4096, hidden_dim=4096, out_dim1=768, use_residual=True): super().__init__() self.layernorm = nn.LayerNorm(in_dim) self.projector = nn.Sequential( nn.Linear(in_dim, hidden_dim, bias=False), nn.GELU(), nn.Linear(hidden_dim, hidden_dim, bias=False), nn.GELU(), nn.Linear(hidden_dim, out_dim, bias=False), ) self.fc = nn.Linear(out_dim, out_dim1) def forward(self, x): x = self.layernorm(x) x = self.projector(x) x2 = nn.GELU()(x) x1 = self.fc(x2) x1 = torch.mean(x1,1) return x1,x2 dtype = torch.bfloat16 device = "cuda" ckpt_id = "black-forest-labs/FLUX.1-schnell" text_encoder_ckpt_id = 'google/byt5-xxl' proj_t5 = MLP(in_dim=4672, out_dim=4096, hidden_dim=4096, out_dim1=768).to(device=device,dtype=dtype) text_encoder_t5 = T5ForConditionalGeneration.from_pretrained(text_encoder_ckpt_id).get_encoder().to(device=device,dtype=dtype) tokenizer_t5 = AutoTokenizer.from_pretrained(text_encoder_ckpt_id) proj_t5_save_path = f"diffusion_pytorch_model.bin" state_dict = torch.load(proj_t5_save_path, map_location="cpu") state_dict_new = {} for k,v in state_dict.items(): k_new = k.replace("module.","") state_dict_new[k_new] = v proj_t5.load_state_dict(state_dict_new) pipeline = FluxPipeline.from_pretrained( ckpt_id, text_encoder=None, text_encoder_2=None, tokenizer=None, tokenizer_2=None, vae=None, torch_dtype=torch.bfloat16 ).to(device) vae = AutoencoderKL.from_pretrained( ckpt_id, subfolder="vae", torch_dtype=torch.bfloat16 ).to(device) vae_scale_factor = 2 ** (len(vae.config.block_out_channels)) image_processor = VaeImageProcessor(vae_scale_factor=vae_scale_factor) while True: raw_text = input("\nPlease Input Query (stop to exit) >>> ") if not raw_text: print('Query should not be empty!') continue if raw_text == "stop": break with torch.no_grad(): text_inputs = tokenizer_t5( raw_text, padding="max_length", max_length=256, truncation=True, add_special_tokens=True, return_tensors="pt", ).input_ids.to(device) text_embeddings = text_encoder_t5(text_inputs)[0] pooled_prompt_embeds,prompt_embeds = proj_t5(text_embeddings) height, width = 1024, 1024 latents = pipeline( prompt_embeds=prompt_embeds, pooled_prompt_embeds=pooled_prompt_embeds, num_inference_steps=4, guidance_scale=0, height=height, width=width, output_type="latent", ).images latents = FluxPipeline._unpack_latents(latents, height, width, vae_scale_factor) latents = (latents / vae.config.scaling_factor) + vae.config.shift_factor image = vae.decode(latents, return_dict=False)[0] image = image_processor.postprocess(image, output_type="pil") image[0].save("MultilingualFLUX.jpg") ``` ## `MultilingualOpenFLUX.1` [OpenFLUX.1] (https://huggingface.co/ostris/OpenFLUX.1) is a fine tune of the FLUX.1-schnell model that has had the distillation trained out of it. Please be sure to update the path of fast-lora.safetensors you have downloaded in the following code. ```python from diffusers import FluxPipeline, AutoencoderKL from diffusers.image_processor import VaeImageProcessor from transformers import T5ForConditionalGeneration,AutoTokenizer import torch import torch.nn as nn class MLP(nn.Module): def __init__(self, in_dim=4096, out_dim=4096, hidden_dim=4096, out_dim1=768, use_residual=True): super().__init__() self.layernorm = nn.LayerNorm(in_dim) self.projector = nn.Sequential( nn.Linear(in_dim, hidden_dim, bias=False), nn.GELU(), nn.Linear(hidden_dim, hidden_dim, bias=False), nn.GELU(), nn.Linear(hidden_dim, out_dim, bias=False), ) self.fc = nn.Linear(out_dim, out_dim1) def forward(self, x): x = self.layernorm(x) x = self.projector(x) x2 = nn.GELU()(x) x1 = self.fc(x2) x1 = torch.mean(x1,1) return x1,x2 dtype = torch.bfloat16 device = "cuda" ckpt_id = "ostris/OpenFLUX.1" text_encoder_ckpt_id = 'google/byt5-xxl' proj_t5 = MLP(in_dim=4672, out_dim=4096, hidden_dim=4096, out_dim1=768).to(device=device,dtype=dtype) text_encoder_t5 = T5ForConditionalGeneration.from_pretrained(text_encoder_ckpt_id).get_encoder().to(device=device,dtype=dtype) tokenizer_t5 = AutoTokenizer.from_pretrained(text_encoder_ckpt_id) proj_t5_save_path = f"diffusion_pytorch_model.bin" state_dict = torch.load(proj_t5_save_path, map_location="cpu") state_dict_new = {} for k,v in state_dict.items(): k_new = k.replace("module.","") state_dict_new[k_new] = v proj_t5.load_state_dict(state_dict_new) pipeline = FluxPipeline.from_pretrained( ckpt_id, text_encoder=None, text_encoder_2=None, tokenizer=None, tokenizer_2=None, vae=None, torch_dtype=torch.bfloat16 ).to(device) pipeline.load_lora_weights("ostris/OpenFLUX.1/openflux1-v0.1.0-fast-lora.safetensors") vae = AutoencoderKL.from_pretrained( ckpt_id, subfolder="vae", torch_dtype=torch.bfloat16 ).to(device) vae_scale_factor = 2 ** (len(vae.config.block_out_channels)) image_processor = VaeImageProcessor(vae_scale_factor=vae_scale_factor) while True: raw_text = input("\nPlease Input Query (stop to exit) >>> ") if not raw_text: print('Query should not be empty!') continue if raw_text == "stop": break with torch.no_grad(): text_inputs = tokenizer_t5( raw_text, padding="max_length", max_length=256, truncation=True, add_special_tokens=True, return_tensors="pt", ).input_ids.to(device) text_embeddings = text_encoder_t5(text_inputs)[0] pooled_prompt_embeds,prompt_embeds = proj_t5(text_embeddings) height, width = 1024, 1024 latents = pipeline( prompt_embeds=prompt_embeds, pooled_prompt_embeds=pooled_prompt_embeds, num_inference_steps=4, guidance_scale=0, height=height, width=width, output_type="latent", ).images latents = FluxPipeline._unpack_latents(latents, height, width, vae_scale_factor) latents = (latents / vae.config.scaling_factor) + vae.config.shift_factor image = vae.decode(latents, return_dict=False)[0] image = image_processor.postprocess(image, output_type="pil") image[0].save("MultilingualOpenFLUX.jpg") ``` To learn more check out the [diffusers](https://huggingface.co/docs/diffusers/main/en/api/pipelines/flux) documentation # License The adapter itself is Apache License 2.0, but it must follow the license of the main model, such as FLUX.1 [dev] Non Commercial License.