Awesome

T5 Textual Inversion for DeepFloyd IF on a 24 GB GPU

T5_Inversion.ipynb is Copyright © 2023 HANS ROETTGER and distributed under the terms of AGPLv3.

This is an implementation of the textual inversion algorithm to incorporate your own objects, faces, logos or styles into DeepFloyd IF. Input: a couple of original images. Output: an T5 embedding for a single token, that can be used in the standard DeepFloyd IF dream pipeline to generate your artefacts. Also available: implementation variant for Stable Diffusion XL.

I know you don't care about copyright, but at least leave me a ⭐ star in the top right!

Original Images ➡ T5 Embedding

<img src="./samples/input.png" alt="" border=3></img>

Run the notebook to generate and save the T5 Embedding for your images. Sample Embeddings

T5 Embedding ➡ Deep Floyd IF Dream Pipeline

Load the T5 Embedding to a single token (e.g. "my") and use it in the standard DeepFloyd IF dream prompts. T5 Embeddings trained with the IF-I-M-v1.0 model work with the IF-I-XL-v1.0 as well!

def set_embedding(t5,emb,word="my"):
    with torch.no_grad():
        tokens=t5.tokenizer(word,max_length=77,padding='max_length',truncation=True,return_attention_mask=False,add_special_tokens=True,return_tensors='pt')
        assert tokens['input_ids'][0][1]==1, 'word is more then one token'
        tokenNo=tokens['input_ids'][0][0]
        assert t5.model.shared.weight[tokenNo].shape==emb.shape, 'wrong dimension of embedding'
        t5.model.shared.weight[tokenNo]=emb.to(t5.device)

def load_embedding(t5,word="my",embedding_file="myToken.pt",no=0,path="./Embeddings/"):
    emb=torch.load(path+embedding_file)
    set_embedding(t5,emb,word)

load_embedding(t5,word="my",embedding_file="myPuppet.pt",path="./Embeddings/")
prompt="a photo of my at the beach"

Better results with more input images and more training steps

The myNN.pt embedding was trained on 33 neural network images

with learning_rates=[(50,0.5),(100,0.4),(200,0.3),(500,0.2),(2000,0.1),(2000,0.05)]

load_embedding(t5,word="it",embedding_file="myNN.pt",path="./Embeddings/")
prompt="a photo of it"

prompt="a photo of a head made of it"

The myMolecule.pt embedding was trained on 200 ball-and-stick model images

with learning_rates=[(50,0.5),(100,0.4),(200,0.3),(500,0.2),(2000,0.1),(2000,0.05)]

load_embedding(t5,word="that",embedding_file="myMolecule.pt",path="./Embeddings/")
prompt="3D rendering of that"

prompt="3D rendering of a robot made of that"

Prerequisites

• A working DeepFloyd IF environment
• A GPU with at least 24 GB CUDA memory

Installation

• Copy the T5_Inversion.ipynb notebook into your DeepFloyd IF environment. All you need is in the small notebook.
• Set the paths to your local models at the top of the notebook. Restart and run all!
• The t5-v1_1-xxl/IF-I-M-v1.0 models in training mode closely fit into 24 GB CUDA memory.
• If you run into out of memory errors, try to remove "safety_checker" and "watermarker" from the "model_index.json" files to save up some memory during training (see template in this repository).

Issues

• Since DeepFloyd IF uses a puny 64x64 pixel resolution in Stage_I, the results are not as good as textual inversion for stable diffusion, but far better than expected.
• I haven't explored all parameter combinations yet - especially the learning rates are tricky. Take this implementation as a starting point for your own experiments.
• If you get totally blurred results, DeepFlod IF has detected "nudity" in /deepfloyd_if/modules/base.py (226) sample = self.__validate_generations(sample).
• I couldn't try out the T5 Inversion with the IF-I-XL-v1.0 model due to CUDA memory restrictions. Maybe there is someone out there with a RTX6000, A100 or H100? Feedback welcome!