Repository navigation
Expand file tree
/
Copy pathNOTICE
More file actions
110 lines (85 loc) · 5.75 KB
/
Copy pathNOTICE
File metadata and controls
110 lines (85 loc) · 5.75 KB
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
ReFlowSET — SAR-to-EO image translation with a rectified flow bridge
Copyright 2026 Jeonghyeok Do, Seungchul Lee, and Munchurl Kim
This product includes software developed by the authors above, licensed under
the Apache License, Version 2.0 (see LICENSE).
The trained model weights are NOT covered by LICENSE. See LICENSE-WEIGHTS.md.
================================================================================
1. Third-party model code (Apache-2.0)
================================================================================
FLUX.2 (Black Forest Labs)
Licensed under the Apache License, Version 2.0
https://huggingface.co/black-forest-labs/FLUX.2-klein-base-4B
a) Transformer primitives. src/reflowset/transformer_reflowset.py adapts
SingleStreamBlock, DoubleStreamBlock, Modulation, QKNorm, RMSNorm,
EmbedND / rope / apply_rope, MLPEmbedder and timestep_embedding from
Black Forest Labs' FLUX.2 model code. MODIFIED: the text stream is
replaced by a SAR stream, the conditioning path and the output head
differ, and the training-only REPA projector is absent from this
inference port.
b) Autoencoder. src/reflowset/autoencoder_flux2.py is a port of the same
upstream's autoencoder, used FROZEN and never fine-tuned. MODIFIED:
re-expressed as a diffusers ModelMixin; the encoder returns the posterior
mean and discards the log-variance; a 2x2 space-to-depth pack and the
shipped BatchNorm latent normaliser are applied at the module boundary.
The autoencoder WEIGHTS are redistributed with the checkpoints. They are the
Apache-2.0 file from FLUX.2-klein-base-4B, re-serialised from diffusers key
names to the upstream layout and cast to bfloat16. Tensors were paired BY
VALUE, not by an assumed rename table; the round trip is bit-identical to
the source through a full encode/decode. scripts/convert_flux2_ae.py plus
assets/flux2_ae_key_map.json reproduce the file from the official download.
SUBSTITUTED AFTER TRAINING, and stated plainly because it matters: both
released arms were TRAINED and EVALUATED with the FLUX.2-dev serialisation
of this autoencoder. That file encoded and decoded every image during
training and during the evaluation that produced the paper's table. It is
published under the FLUX Non-Commercial License, whose Section 4(a)(iii)
forbids "surveillance purposes, including all research and development
related to surveillance" and whose Section 1(a) makes the restriction
inherit permanently.
The file published here is the Apache-2.0 serialisation of the SAME network,
from FLUX.2-klein-base-4B. 250 of 251 tensors pair one-to-one by value with
a worst absolute deviation of 7.80e-03 (bfloat16 rounding; 3.89e-03
relative), and the unpaired entry is a BatchNorm step counter rather than a
weight. Swapping it into the released checkpoints, measured end to end,
changes QXS-SAROPT PSNR by less than 0.004 dB in absolute value; four
independent measurements at different guidance scales and sample sets land
between -0.004 and +0.002 dB, so the sign is not resolved and only the
magnitude is meaningful. For scale, changing only the evaluation seed moves
the same measurement by +0.395 dB. Nothing the paper prints changes.
Apache-2.0 Section 6 withholds trademark rights. This model is named
ReFlowSET and must not be presented as a FLUX product or endorsed by
Black Forest Labs.
Hugging Face Diffusers
Licensed under the Apache License, Version 2.0
https://github.com/huggingface/diffusers
Pipeline, scheduler and model base classes and their conventions.
================================================================================
2. Third-party components used in TRAINING and NOT redistributed
================================================================================
DINOv3 (Meta Platforms) — frozen REPA teacher, ViT-L/16 LVD-1689M
Used during training only, with lambda_repa = 0.5, to align block-8
activations with the teacher's patch tokens. The teacher is not loaded at
inference and the projector that consumed it is not part of the released
checkpoints.
NO DINOv3 WEIGHTS ARE REDISTRIBUTED HERE. Obtain them from Meta's official
release and accept the DINOv3 License:
https://github.com/facebookresearch/dinov3
DINOv3 License Section 1(b)(ii) requires that publications reporting research
performed using DINO Materials acknowledge that use; this NOTICE and the
paper's acknowledgements do so. Section 1(b)(v) restricts use under Trade
Controls, including military, warfare and espionage end uses; anyone
retraining ReFlowSET with this teacher is subject to those terms directly.
================================================================================
3. Datasets — used, not redistributed
================================================================================
QXS-SAROPT https://github.com/yaoxu008/QXS-SAROPT
SAR2Opt https://github.com/MarsZhaoYT/SAR2Opt-Heterogeneous-Dataset
Neither dataset is redistributed by this repository. The frozen split lists in
splits/ name files within those datasets and contain no imagery.
================================================================================
4. Comparison methods
================================================================================
The fifteen prior methods reproduced for the paper's main table were each trained
with their own upstream implementation, and every resulting checkpoint inherits
that implementation's terms. Four of them ship no upstream LICENSE file at all.
Per-method attribution, licence text and the exact wording each one requires are
in MODEL_ZOO.md and in the model cards accompanying the released weights.