loralover/xv7k-fm-weights-archive

Flow Matching Checkpoints (CelebA-64)

Model checkpoints from "From Diffusion to One-Step Generation: A Controlled Study of Flow Matching, Rectified Flow, and Inference Acceleration on CelebA-64".

Code: github.com/kshitiz-1225/flowmatch-code

Architecture

• Model: UNet (77.7M parameters)
• Dataset: CelebA-64×64
• Framework: PyTorch

Checkpoint Index

Teacher

File	Description
teacher/2rf_rfpp.pt	2-rectified-flow teacher (RFPP), used to generate all reflow pairs

Core Ablations (Reflow Factorial Study)

Experiment	Config	Checkpoints	KID@1 (×1e-3)
ablation_A_uniform_mse	Uniform + MSE	final, 50k, 100k	8.01
ablation_B_ushaped_mse	U-shaped + MSE	final, 50k, 100k	8.19
ablation_C_uniform_lpips_huber	Uniform + LPIPS-Huber	final, 50k, 100k	~7.76
ablation_D_ushaped_lpips_huber	U-shaped + LPIPS-Huber	final, 50k, 100k	8.21

Extended Ablations

Experiment	Config	Checkpoints
ext_B2_beta2_mse	Beta(2,2) + MSE	final, 50k, 100k
ext_Bfix_rfpp_tdist_mse	RFPP t-dist + MSE	final, 50k, 100k
ext_Cext_uniform_lpips_huber_200k	Uniform + LPIPS-Huber (200k iters)	final, 150k, 200k
ext_D2_beta2_lpips_huber	Beta(2,2) + LPIPS-Huber	final, 100k, 200k
ext_Dext_ushaped_lpips_huber_200k	U-shaped + LPIPS-Huber (200k)	final, 150k, 200k
ext_Dfix_rfpp_tdist_lpips_huber	RFPP t-dist + LPIPS-Huber	final, 100k, 200k

Acceleration Methods

Experiment	Method	KID@1 (×1e-3)	KID@2	KID@5
ext_I_ect	Easy Consistency Tuning	6.56	5.76	5.43
ext_J_consistency_fm	Consistency Flow Matching	6.75	5.30	4.71
ext_K_pcm	Phased Consistency Model	6.64	5.98	5.60
ext_M_meanflow	MeanFlow	9.02	6.49	5.53
ext_G_shortcut	Shortcut conditioning	415.5	—	—
ext_F_self_distill	Self-distillation	92.3	—	—
ext_L_gan_distill	GAN distillation	~20	—	—

Timestep Distribution Variants

Experiment	Description
ext_H_adaptive_t	Learned adaptive timestep schedule
ext_H_uniformhist_lpips	Uniform histogram + LPIPS
ext_H_mixhist_lpips	Mixed histogram + LPIPS
ext_H_rfpphist_lpips	RFPP-shaped histogram + LPIPS
ext_H_adaptive_t_importance_weighted	Importance-weighted adaptive
ext_H_adaptive_t_objective_shaping	Objective-shaped LPIPS

Distillation Variants

Experiment	Description
ext_F_preempt_s4	Progressive distill, stride=4
ext_F_preempt_s8	Progressive distill, stride=8
ext_F_preempt_s8_lp02	Progressive distill, stride=8, LPIPS=0.2
ext_F_preempt_s16	Progressive distill, stride=16

Baseline

Experiment	Description
hw3_reflow_baseline	Flow matching teacher → reflow baseline (uniform+MSE, 100k)

Usage

import torch
from huggingface_hub import hf_hub_download

# Download a specific checkpoint
path = hf_hub_download(
    repo_id="loralover/xv7k-fm-weights-archive",
    filename="archive/ablation_A_uniform_mse/reflow_final.pt",
)
checkpoint = torch.load(path, map_location="cpu")

Directory Structure

teacher/                    # 2-rectified-flow teacher
archive/<experiment>/       # From checkpoint archive (3 ckpts each)
logs/<experiment>/          # From training logs (1-3 ckpts each)

Citation

See the code repository for full details.