File size: 3,782 Bytes
0161e74 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 | ## Latent Forcing: Reordering the Diffusion Trajectory for Pixel-Space Image Generation
[](https://arxiv.org/abs/2602.11401)
<!-- <p align="center">
<img src="demo/visual.jpg" width="100%">
</p> -->
Code for Latent Forcing
```
@article{baade2026latentforcing,
title={Latent Forcing: Reordering the Diffusion Trajectory for Pixel-Space Image Generation},
author={Alan Baade and Eric Ryan Chan and Kyle Sargent and Changan Chen and Justin Johnson and Ehsan Adeli and Li Fei-Fei},
journal={arXiv preprint arXiv:2602.11401},
year={2026},
}
```
Our code is based on JiT: https://github.com/LTH14/JiT.git
<p align="left">
<img src="assets/ConceptDiagram.png" width="40%">
</p>
### Dataset
We use [ImageNet](http://image-net.org/download) dataset, and Webdataset.
### Installation
Download the code:
```
git clone https://github.com/AlanBaade/LatentForcing.git
cd LatentForcing
```
Create the conda environment. uv is recommended, but not required.
```bash
conda create -n latentforcing python=3.10
conda activate latentforcing
uv pip install opencv-python==4.11.0.86 numpy==1.23 timm==0.9.12 tensorboard==2.10.0 scipy==1.9.1 einops==0.8.1 gdown==5.2.0 matplotlib==3.10.8 transformers==4.57.3 webdataset==1.0.2
uv pip install torch==2.5.1 --index-url https://download.pytorch.org/whl/cu124
uv pip install "torch-fidelity @ git+https://github.com/LTH14/torch-fidelity.git@master"
```
### Training
Example script for training LatentForcing-L on ImageNet 200 epochs:
```
torchrun --nproc_per_node=8 --standalone \
main_jit.py \
--model JiTCoT-LM/16 \
--D_mean -1.2 --D_std 1.0 \
--P_mean -0.4 --P_std 0.8 \
--batch_size 128 --blr 5e-5 \
--epochs 200 --warmup_epochs 5 \
--gen_bsz 256 --num_images 10000 \
--cfg 1.0 --cfg_dino 1.0 \
--interval_min 0.0 --interval_max 1.0 \
--dino_weight 0.333 --choose_dino_p 0.4 \
--sample_mode dino_first_cascaded_noised \
--dh_depth 2 --dh_hidden_size 1024 \
--output_dir ${OUTPUT_DIR} \
--resume ${OUTPUT_DIR} \
--data_path ${DATA_PATH} \
--online_eval
```
For unconditional training and generation, set ```--label_drop_prob 1.0```
To train a Multi-Schedule model set ```--sample_mode shifted_independent_uniform```
### Evaluation
PyTorch pre-trained models are WIP
Evaluate LatentForcing-L with Autoguidance (Default Evaluation Setting)
```
torchrun --nproc_per_node=8 --standalone \
main_jit.py \
--model JiTCoT-LM/16 \
--dh_depth 2 --dh_hidden_size 1024 \
--gen_bsz 1536 --num_images 50000 \
--cfg 1.5 --cfg_dino 1.5 \
--interval_min 0.0 --interval_max 1.0 \
--interval_min_dino 0.0 --interval_max_dino 1.0 \
--sample_mode dino_first_cascaded_noised \
--output_dir ${OUTPUT_DIR_EVAL} \
--resume ${OUTPUT_DIR} \
--data_path ${DATA_PATH} \
--evaluate_gen --num_sampling_steps 50 \
--sampling_method heun \
--guidance_method autoguidance \
--autoguidance_ckpt ${AUTOGUIDANCE_CKPT}$
```
Evaluate LatentForcing-L with Interval CFG (Used in the System-Level comparison only)
```
torchrun --nproc_per_node=8 --standalone \
main_jit.py \
--model JiTCoT-LM/16 \
--dh_depth 2 --dh_hidden_size 1024 \
--gen_bsz 1536 --num_images 50000 \
--cfg 1.5 --cfg_dino 2.9 \
--interval_min 0.0 --interval_max 1.0 \
--interval_min_dino 0.06 --interval_max_dino 1.0 \
--sample_mode dino_first_cascaded_noised \
--output_dir ${OUTPUT_DIR_EVAL} \
--resume ${OUTPUT_DIR} \
--data_path ${DATA_PATH} \
--evaluate_gen --num_sampling_steps 50 \
--gen_shift_dino 0.575 --sampling_method heun \
--guidance_method cfg_interval \
--autoguidance_ckpt ${AUTOGUIDANCE_CKPT}$
```
We use the same customized FID eval as JiT: [```torch-fidelity```](https://github.com/LTH14/torch-fidelity)
### Contact
You can contact me at baade@stanford.edu for questions.
|