EDM2-diffusers
Diffusers-ready checkpoints for
EDM2 (
Analyzing and Improving the Training Dynamics of Diffusion Models ),
converted from
NVlabs/edm2 post-hoc reconstructions.
Official source weights: https://nvlabs-fi-cdn.nvidia.com/edm2/posthoc-reconstructions/
This root folder is a model collection that contains:
edm2-img512-xs-fidedm2-img512-s-fidedm2-img512-m-fidedm2-img512-l-fidedm2-img512-l-dinoedm2-img512-xl-fidedm2-img512-xxl-fid
Each subfolder is a self-contained Diffusers model repo with:
pipeline.pyunet/unet_edm2.pyscheduler/scheduler_config.json (EDMEulerScheduler)unet/diffusion_pytorch_model.safetensorsvae/diffusion_pytorch_model.safetensors
Demo
edm2-img512-xxl-fid demo Class-conditional sample (ImageNet class 207 , golden retriever), EDM2-XXL at 512×512, 32 steps, guidance 1.0, seed 42.
Model Paths
Use paths relative to this root README:
Model NVlabs preset FID Local path EDM2-XS edm2-img512-xs-fid3.53 ./edm2-img512-xs-fidEDM2-S edm2-img512-s-fid2.56 ./edm2-img512-s-fidEDM2-M edm2-img512-m-fid2.25 ./edm2-img512-m-fidEDM2-L edm2-img512-l-fid2.06 ./edm2-img512-l-fidEDM2-L (DINO) edm2-img512-l-dino— ./edm2-img512-l-dinoEDM2-XL edm2-img512-xl-fid1.96 ./edm2-img512-xl-fidEDM2-XXL edm2-img512-xxl-fid1.91 ./edm2-img512-xxl-fid
Inference Demo (Diffusers)
1) Load a local subfolder checkpoint
1 from pathlib import Path
2 import torch
3 from diffusers import DiffusionPipeline
4
5 model_dir = Path ( "./edm2-img512-xxl-fid" ) # change to any path in the table above
6 pipe = DiffusionPipeline . from_pretrained (
7 str ( model_dir ) ,
8 local_files_only = True ,
9 trust_remote_code = True ,
10 torch_dtype = torch . bfloat16 ,
11 ) . to ( "cuda" )
12
13 generator = torch . Generator ( device = "cuda" ) . manual_seed ( 42 )
14 image = pipe (
15 class_labels = 207 , # golden retriever (ImageNet id); omit for random class
16 num_inference_steps = 32 ,
17 guidance_scale = 1.0 , # >1.0 requires a gnet/ checkpoint
18 generator = generator ,
19 ) . images [ 0 ]
20 image . save ( "demo.png" )
Official inference defaults (generate_images.py): num_steps=32, sigma_min=0.002,
sigma_max=80, rho=7, guidance=1.0 (no gnet), S_churn=0. Heun sampling runs in
float32 internally even when UNet/VAE weights are loaded in bf16/fp16.
Guided presets require a converted gnet/ folder and guidance_scale matching the
NVlabs preset.
2) Convert a legacy .pkl
1 python scripts/convert_edm2_to_diffusers.py \
2 --checkpoint models/BiliSakura/EDM2-diffusers/edm2-img512-xs-2147483-0.135.pkl \
3 --output models/BiliSakura/EDM2-diffusers
Creates edm2-img512-xs-fid/ automatically from the NVlabs preset mapping.
Checkpoint preset mapping
Maps NVlabs
--preset=... names from
generate_images.py
to source pickle filenames and local Diffusers directories.
EDM2 paper — ImageNet-512 (conditional)
NVlabs preset Source .pkl (net) Diffusers dir Metric edm2-img512-xs-fidedm2-img512-xs-2147483-0.135.pkledm2-img512-xs-fid/FID 3.53 edm2-img512-xs-dinoedm2-img512-xs-2147483-0.200.pkl— FDDINOv2 103.39 edm2-img512-s-fidedm2-img512-s-2147483-0.130.pkledm2-img512-s-fid/FID 2.56 edm2-img512-s-dinoedm2-img512-s-2147483-0.190.pkl— FDDINOv2 68.64 edm2-img512-m-fidedm2-img512-m-2147483-0.100.pkledm2-img512-m-fid/FID 2.25 edm2-img512-m-dinoedm2-img512-m-2147483-0.155.pkl— FDDINOv2 58.44 edm2-img512-l-fidedm2-img512-l-1879048-0.085.pkledm2-img512-l-fid/FID 2.06 edm2-img512-l-dinoedm2-img512-l-1879048-0.155.pkledm2-img512-l-dino/FDDINOv2 52.25 edm2-img512-xl-fidedm2-img512-xl-1342177-0.085.pkledm2-img512-xl-fid/FID 1.96 edm2-img512-xl-dinoedm2-img512-xl-1342177-0.155.pkl— FDDINOv2 45.96 edm2-img512-xxl-fidedm2-img512-xxl-0939524-0.070.pkledm2-img512-xxl-fid/FID 1.91 edm2-img512-xxl-dinoedm2-img512-xxl-0939524-0.150.pkl— FDDINOv2 42.84
EDM2 paper — ImageNet-64 (conditional)
NVlabs preset Source .pkl (net) Metric edm2-img64-s-fidedm2-img64-s-1073741-0.075.pklFID 1.58 edm2-img64-m-fidedm2-img64-m-2147483-0.060.pklFID 1.43 edm2-img64-l-fidedm2-img64-l-1073741-0.040.pklFID 1.33 edm2-img64-xl-fidedm2-img64-xl-0671088-0.040.pklFID 1.33
EDM2 paper — classifier-free guidance (ImageNet-512)
Use guidance_scale below and include the converted gnet/ checkpoint.
NVlabs preset Source .pkl (net) Source .pkl (gnet) Guidance Metric edm2-img512-xs-guid-fidedm2-img512-xs-2147483-0.045.pkledm2-img512-xs-uncond-2147483-0.045.pkl1.40 FID 2.91 edm2-img512-xs-guid-dinoedm2-img512-xs-2147483-0.150.pkledm2-img512-xs-uncond-2147483-0.150.pkl1.70 FDDINOv2 79.94 edm2-img512-s-guid-fidedm2-img512-s-2147483-0.025.pkledm2-img512-xs-uncond-2147483-0.025.pkl1.40 FID 2.23 edm2-img512-s-guid-dinoedm2-img512-s-2147483-0.085.pkledm2-img512-xs-uncond-2147483-0.085.pkl1.90 FDDINOv2 52.32 edm2-img512-m-guid-fidedm2-img512-m-2147483-0.030.pkledm2-img512-xs-uncond-2147483-0.030.pkl1.20 FID 2.01 edm2-img512-m-guid-dinoedm2-img512-m-2147483-0.015.pkledm2-img512-xs-uncond-2147483-0.015.pkl2.00 FDDINOv2 41.98 edm2-img512-l-guid-fidedm2-img512-l-1879048-0.015.pkledm2-img512-xs-uncond-2147483-0.015.pkl1.20 FID 1.88 edm2-img512-l-guid-dinoedm2-img512-l-1879048-0.035.pkledm2-img512-xs-uncond-2147483-0.035.pkl1.70 FDDINOv2 38.20 edm2-img512-xl-guid-fidedm2-img512-xl-1342177-0.020.pkledm2-img512-xs-uncond-2147483-0.020.pkl1.20 FID 1.85 edm2-img512-xl-guid-dinoedm2-img512-xl-1342177-0.030.pkledm2-img512-xs-uncond-2147483-0.030.pkl1.70 FDDINOv2 35.67 edm2-img512-xxl-guid-fidedm2-img512-xxl-0939524-0.015.pkledm2-img512-xs-uncond-2147483-0.015.pkl1.20 FID 1.81 edm2-img512-xxl-guid-dinoedm2-img512-xxl-0939524-0.015.pkledm2-img512-xs-uncond-2147483-0.015.pkl1.70 FDDINOv2 33.09
Autoguidance paper
NVlabs preset Source .pkl (net) Source .pkl (gnet) Guidance Metric edm2-img512-s-autog-fidedm2-img512-s-2147483-0.070.pkledm2-img512-xs-0134217-0.125.pkl2.10 FID 1.34 edm2-img512-s-autog-dinoedm2-img512-s-2147483-0.120.pkledm2-img512-xs-0134217-0.165.pkl2.45 FDDINOv2 36.67 edm2-img512-xxl-autog-fidedm2-img512-xxl-0939524-0.075.pkledm2-img512-m-0268435-0.155.pkl2.05 FID 1.25 edm2-img512-xxl-autog-dinoedm2-img512-xxl-0939524-0.130.pkledm2-img512-m-0268435-0.205.pkl2.30 FDDINOv2 24.18 edm2-img512-s-uncond-autog-fidedm2-img512-s-uncond-2147483-0.070.pkledm2-img512-xs-uncond-0134217-0.110.pkl2.85 FID 3.86 edm2-img512-s-uncond-autog-dinoedm2-img512-s-uncond-2147483-0.090.pkledm2-img512-xs-uncond-0134217-0.125.pkl2.90 FDDINOv2 90.39 edm2-img64-s-autog-fidedm2-img64-s-1073741-0.045.pkledm2-img64-xs-0134217-0.110.pkl1.70 FID 1.01 edm2-img64-s-autog-dinoedm2-img64-s-1073741-0.105.pkledm2-img64-xs-0134217-0.175.pkl2.20 FDDINOv2 31.85
NVlabs preset shorthand
1 # EDM2 paper
2 edm2-img512-{xs|s|m|l|xl|xxl}-{fid|dino}
3 edm2-img64-{s|m|l|xl}-fid
4 edm2-img512-{xs|s|m|l|xl|xxl}-guid-{fid|dino}
5
6 # Autoguidance paper
7 edm2-img512-{s|xxl}-autog-{fid|dino}
8 edm2-img512-s-uncond-autog-{fid|dino}
9 edm2-img64-s-autog-{fid|dino}
Example NVlabs command:
python generate_images.py --preset=edm2-img512-s-guid-dino --outdir=outEquivalent expanded form:
1 python generate_images.py \
2 --net = https://nvlabs-fi-cdn.nvidia.com/edm2/posthoc-reconstructions/edm2-img512-s-2147483-0.085.pkl \
3 --gnet = https://nvlabs-fi-cdn.nvidia.com/edm2/posthoc-reconstructions/edm2-img512-xs-uncond-2147483-0.085.pkl \
4 --guidance = 1.9 \
5 --outdir = out
