tensorkelechi/arxiv-cstext · Datasets at Hugging Face

text stringlengths 0 89.3k
35 Sean Welleck and Rahul Saha Llmstep Llm proofstep suggestions in lean arXiv preprint
arXiv231018457 2023
36 Sean Welleck Neural theorem proving tutorial ii httpsgithubcomcmul3
ntptutorialII 2023
A Appendix
This is an incorrect proof of the heap sort orders list elements property GPT4 leaves in placeholders
as arguments to lemmas from the environment that it does not use in the proof It also attempts the
invalid operation of induction generalizing the inductive type MyHeap Transforming this proof
into a correct one is nontrivial
unfold hsort
unfold ordered
induction xs generalizing MyHeap with
nil unfold toHeap unfold toList simp
cons x xs ih
unfold toHeap
unfold toList
unfold hmerge
have h numElem_merge_branches_lt Placeholder for specific heaps if
needed
have m merge_elems Placeholder for specific heaps if needed
11specialize ih xsmap fun x MyHeapnode MyHeapnil x MyHeapnil
induction toList toHeap xs generalizing toList toHeap x xs
nil
apply ih
cons y ys ysih
unfold toList at ih
unfold toHeap at ih
unfold ordered at ih
unfold hsort at ih
have h numElem_merge_branches_lt Placeholder for specific values if
needed
have m merge_elems Placeholder for specific values if needed
case h_1
unfold ordered
simp Listmap
simp toList
case h_2
unfold toHeap
unfold ordered
simp toList
rfl
12ConsistDreamer 3DConsistent 2D Diffusion for HighFidelity Scene Editing
JunKun Chen1Samuel Rota Bul o2Norman M uller2Lorenzo Porzi2
Peter Kontschieder2YuXiong Wang1
1University of Illinois UrbanaChampaign2Meta
junkun3yxw illinoisedu
rotabulonormanmporzipkontschieder metacom
Figure 1 Our ConsistDreamer lifts 2D diffusion with 3D awareness and consistency achieving highfidelity instructionguided scene
editing with superior sharpness and detailed textures Left The three synergistic components within ConsistDreamer that enable 3D
consistency Right Stateoftheart performance of ConsistDreamer across various editing tasks and scenes especially when prior work
eg IN2N 8 fails and in challenging largescale indoor scenes from ScanNet 43 More results are on our project page
Abstract
This paper proposes ConsistDreamer a novel frame
work that lifts 2D diffusion models with 3D awareness
and 3D consistency thus enabling highfidelity instruction
guided scene editing To overcome the fundamental limi
tation of missing 3D consistency in 2D diffusion models
our key insight is to introduce three synergistic strategies
that augment the input of the 2D diffusion model to be
come 3Daware and to explicitly enforce 3D consistency
during the training process Specifically we design sur
rounding views as contextrich input for the 2D diffusion
model and generate 3Dconsistent structured noise instead
of imageindependent noise Moreover we introduce self
supervised consistencyenforcing training within the per
scene editing procedure Extensive evaluation shows that
our ConsistDreamer achieves stateoftheart performance
for instructionguided scene editing across various scenes
and editing instructions particularly in complicated large
scale indoor scenes from ScanNet with significantly im
proved sharpness and finegrained textures Notably Con
sistDreamer stands as the first work capable of success
fully editing complex eg plaidcheckered patterns Our
project page is at immortalcogithubioConsistDreamer
Work started during an internship at Meta Reality Labs Zurich1 Introduction
With the emergence of instructionguided 2D generative
models as in 2 it has never been easier to generate or
edit images Extending this success to 3D ie instruction
guided 3D scene editing becomes highly desirable for
artists designers and the movie and game industries Nev
ertheless editing 3D scenes or objects is inherently chal
lenging The absence of largescale general 3D datasets
makes it difficult to create a counterpart generative model
similar to 2 that can support arbitrary 3D scenes There
fore stateoftheart solutions 8 25 circumvent this chal
lenge by resorting to generalizable 2D diffusion models
This approach known as 2D diffusion distillation renders
the scene into multiview images applies an instruction
conditioned diffusion model in 2D and then distills the edit
ing signal back to 3D such as through a neural radiance
field NeRF 6 8 15
However a fundamental limitation of this solution is the
lack of 3D consistency a 2D diffusion model acting in
dependently across views is likely to produce inconsistent
edits both in color and shape For example a person in
one view might be edited to be wearing a red shirt while
appearing in a green shirt in another view Using these im
ages to train a NeRF can still produce reasonable edits but
the model will naturally converge towards an averaged
representation of the inconsistent 2D supervision and lose

35 Sean Welleck and Rahul Saha Llmstep Llm proofstep suggestions in lean arXiv preprint

arXiv231018457 2023

36 Sean Welleck Neural theorem proving tutorial ii httpsgithubcomcmul3

ntptutorialII 2023

A Appendix

This is an incorrect proof of the heap sort orders list elements property GPT4 leaves in placeholders

as arguments to lemmas from the environment that it does not use in the proof It also attempts the

invalid operation of induction generalizing the inductive type MyHeap Transforming this proof

into a correct one is nontrivial

unfold hsort

unfold ordered

induction xs generalizing MyHeap with

nil unfold toHeap unfold toList simp

cons x xs ih

unfold toHeap

unfold toList

unfold hmerge

have h numElem_merge_branches_lt Placeholder for specific heaps if

needed

have m merge_elems Placeholder for specific heaps if needed

11specialize ih xsmap fun x MyHeapnode MyHeapnil x MyHeapnil

induction toList toHeap xs generalizing toList toHeap x xs

nil

apply ih

cons y ys ysih

unfold toList at ih

unfold toHeap at ih

unfold ordered at ih

unfold hsort at ih

have h numElem_merge_branches_lt Placeholder for specific values if

needed

have m merge_elems Placeholder for specific values if needed

case h_1

unfold ordered

simp Listmap

simp toList

case h_2

unfold toHeap

unfold ordered

simp toList

rfl

12ConsistDreamer 3DConsistent 2D Diffusion for HighFidelity Scene Editing

JunKun Chen1Samuel Rota Bul o2Norman M uller2Lorenzo Porzi2

Peter Kontschieder2YuXiong Wang1

1University of Illinois UrbanaChampaign2Meta

junkun3yxw illinoisedu

rotabulonormanmporzipkontschieder metacom

Figure 1 Our ConsistDreamer lifts 2D diffusion with 3D awareness and consistency achieving highfidelity instructionguided scene

editing with superior sharpness and detailed textures Left The three synergistic components within ConsistDreamer that enable 3D

consistency Right Stateoftheart performance of ConsistDreamer across various editing tasks and scenes especially when prior work

eg IN2N 8 fails and in challenging largescale indoor scenes from ScanNet 43 More results are on our project page

Abstract

This paper proposes ConsistDreamer a novel frame

work that lifts 2D diffusion models with 3D awareness

and 3D consistency thus enabling highfidelity instruction

guided scene editing To overcome the fundamental limi

tation of missing 3D consistency in 2D diffusion models

our key insight is to introduce three synergistic strategies

that augment the input of the 2D diffusion model to be

come 3Daware and to explicitly enforce 3D consistency

during the training process Specifically we design sur

rounding views as contextrich input for the 2D diffusion

model and generate 3Dconsistent structured noise instead

of imageindependent noise Moreover we introduce self

supervised consistencyenforcing training within the per

scene editing procedure Extensive evaluation shows that

our ConsistDreamer achieves stateoftheart performance

for instructionguided scene editing across various scenes

and editing instructions particularly in complicated large

scale indoor scenes from ScanNet with significantly im

proved sharpness and finegrained textures Notably Con

sistDreamer stands as the first work capable of success

fully editing complex eg plaidcheckered patterns Our

project page is at immortalcogithubioConsistDreamer

Work started during an internship at Meta Reality Labs Zurich1 Introduction

With the emergence of instructionguided 2D generative

models as in 2 it has never been easier to generate or

edit images Extending this success to 3D ie instruction

guided 3D scene editing becomes highly desirable for

artists designers and the movie and game industries Nev

ertheless editing 3D scenes or objects is inherently chal

lenging The absence of largescale general 3D datasets

makes it difficult to create a counterpart generative model

similar to 2 that can support arbitrary 3D scenes There

fore stateoftheart solutions 8 25 circumvent this chal

lenge by resorting to generalizable 2D diffusion models

This approach known as 2D diffusion distillation renders

the scene into multiview images applies an instruction

conditioned diffusion model in 2D and then distills the edit

ing signal back to 3D such as through a neural radiance

field NeRF 6 8 15

However a fundamental limitation of this solution is the

lack of 3D consistency a 2D diffusion model acting in

dependently across views is likely to produce inconsistent

edits both in color and shape For example a person in

one view might be edited to be wearing a red shirt while

appearing in a green shirt in another view Using these im

ages to train a NeRF can still produce reasonable edits but

the model will naturally converge towards an averaged

representation of the inconsistent 2D supervision and lose