EG-FM-ImageNet / README.md
ysng's picture
Update README.md
52d7c63 verified
|
Raw
History Blame Contribute Delete
3.22 kB
metadata
license: cc-by-nc-4.0
library_name: pytorch
pipeline_tag: unconditional-image-generation
datasets:
  - ILSVRC/imagenet-1k
tags:
  - image-generation
  - class-conditional-image-generation
  - flow-matching
  - pixel-space
  - imagenet
arxiv: '2608.05811'
extra_gated_heading: EG-FM Model Weights Access Application
extra_gated_description: >-
  Please complete this form to request access to the EG-FM model weights.
  Requests are reviewed manually.
extra_gated_button_content: Submit access request
extra_gated_prompt: >-
  Thank you for your interest in EG-FM. To support responsible use of the model
  weights, please provide accurate information below. Our team will normally
  review your request within 1 business days. Using an institutional, academic,
  or corporate email address may help us verify your application.
extra_gated_fields:
  Name: text
  Affiliation (Organization / Company / University): text
  Your role:
    type: select
    options:
      - Researcher / Professor
      - Student
      - Engineer / Architect
      - Independent Developer
      - Other
  Intended use case: text
  I confirm that my use is non-commercial and complies with the CC BY-NC license terms: checkbox
  I commit not to redistribute or resell the model weights to any third party: checkbox
  I commit not to use this model for illegal, harmful, or unethical activities: checkbox

Energy-Guided Flow Matching (EG-FM)

Official checkpoints for Energy-Guided Flow Matching (EG-FM) on class-conditional ImageNet-1K generation at 256×256 and 512×512 resolutions.

EG-FM introduces an image-specific, energy-guided moving endpoint to construct an explicit coarse-to-fine generation trajectory while requiring no changes to the backbone or training data.

Released models

Model Resolution Training FID
PixelDiT-200 256×256 200 epochs 1.55
PixelDiT-600 256×256 600 epochs 1.45
PixelDiT-220 512×512 200+20 epochs 1.72
PixelDiT-240 512×512 200+40 epochs 1.68

All FID values use the ADM evaluation suite. The 512×512 models continue from the 200-epoch 256×256 checkpoint; 200+20 and 200+40 denote the initial training followed by additional high-resolution adaptation epochs.

For installation, checkpoint loading, inference, and evaluation commands, see the official GitHub repository.

License

The model weights are released under CC BY-NC 4.0. Commercial use is not permitted under this license.

Citation

@article{tong2026energy,
  title   = {Energy-Guided Flow Matching},
  author  = {Tong, Haoyang and He, Yu and Li, Fang and Ma, Lichen and Fu, Jingling and Chen, Dong and Chen, Zhen and Huang, Junshi and Cao, Jie},
  journal = {arXiv preprint arXiv:2608.05811},
  year    = {2026}
}