text
stringlengths
0
89.3k
Full All 2740415982162719132812
No Str Noise SN 3500523482483122217820
No Training T 3461122182080420201819
Only Sur Views SNT 3400326202783922214420
IN2N SNSVT9101225582090915222420
Table 1 Ablation study on the distillation fidelity score quan
titatively validates the effectiveness and complementarity of each
of our components Our full ConsistDreamer significantly outper
forms all variants across various scenes and types of instructions
with the training objective of DreamFusion 25 we aim to
minimize the distance between two distributions the distri
bution of a rendered image at a random view from the edited
NeRF and the distribution of the diffusion editing result of
an image at a random view in the original scene Following
this we define the fidelity metric as the Fr echet inception
distance FID 9 30 between two sets the set of images
rendered by the edited NeRF at all training views and the
set of edited images generated by the original 2 for all
training views corresponding to these two distributions A
lower FID means a higher fidelity that the editing is applied
to the scene
Qualitative Results The qualitative comparison in the
Fangzhou scene from the IN2N dataset is shown in Fig 3
Distilling from the same diffusion model 2 IN2N 8
ViCA 6 and our ConsistDreamer produce results in a sim
ilar style As especially shown in the Vincent Van Gogh
and Edvard Munch editing our ConsistDreamer gener
ates results containing finegrained representative textures
of Van Gogh and Munch while the baseline results are
blurred and only contain simple or coarse textures This val
idates that with our proposed components ConsistDreamer
is able to generate consistent images from 2 with detailed
textures and does not rely on consistency derived from
NeRF which unfortunately smooths out the results No
tably in the Lord V oldemort case our ConsistDreamer is
theonly one that successfully edits the image to resemble
the wellknown distinctive appearance of Lord V oldemort
featuring no hair and a peculiar nose Among all the editing
tasks our ConsistDreamer consistently produces editing re
sults with the most detailed ears and hairhead and does not
contain unnatural color blocks
Additional qualitative results are shown in Fig 4 and
more results and the comparison with baselines on these
tasks are provided in the supplementary and on our project
page Overall our ConsistDreamer generates sharp bright
editing results in all tasks across various scenes including
human indoor and outdoor scenes 1 In the Face scene
our ConsistDreamer successfully applies the plaid check
ered jacket editing a common failure case in most previous
7Figure 4 ConsistDreamer consistently generates highquality and highfidelity editing results featuring detailed finegrained textures
across various scenes and instructions Notably ConsistDreamer also maintains the high diversity from 2 as exemplified by the highly
diversified results ab Additional results and comparisons are provided in the supplementary and on our project page
methods including IN2N Also our ConsistDreamer is able
to assign finegrained marble texture in the marble statue
editing a clear mustache in Mustache editing and clear
wrinkle and hair in Einstein editing while IN2N produces
blurred and oversmooth results with poor details Notably
our ConsistDreamer minimizes the side effects of the edit
ing while IN2N unexpectedly and significantly changes the
skin color in the Mustache editing and the wall color in the
Einstein editing 2 The Tolkien Elf and Fauvism editing
tasks in the Fangzhou scene show that our ConsistDreamer
could preserve most diversity from the original 2 due
to the use of structured noise sampled for the whole edit
ing With the structured noise we can focus on the con
sistency of generation for the given noise without suffer
ing from averaging results generated from different noises
which may lose diversity by converging to an average style
for all noises 3 Our ConsistDreamer works well in out
door scenes as all the details on the floor mountain plants
and camps are preserved in the edited results 4 In compli
cated indoor scenes from the ScanNet dataset our Con
sistDreamer generates editing results that are easy to rec
ognize as the given style with finegrained textures Van
Gogh regular patterns Picasso or special lighting condi
tions Bastion and Transistor All these results validate that
our ConsistDreamer generates highquality editing results
Ablation Study As shown in Table 1 we conduct the ab
lation study on four representative tasks AD covering
instructions of objectspecific editing artistic style trans
fer and other style transfer and scenes of human indoor
and outdoor scenes The results show that our full ConsistDreamer outperforms all the variants with significant
gains in all tasks under DFS which mainly comes from
our consistent denoising procedure in Sec 33 that requires
all three major components to achieve Training towards a
consistent denoising procedure produces considerable extra
supervision signals to the augmented 2 making it con
verge better towards consistent generation results We can
also observe that the consistencyenforcing training and the
use of surrounding views improve the fidelity in most of
the tasks especially in the complicated largescale indoor
scenes CD showing that these components indeed im
prove the consistency in generation
5 Conclusion
This paper proposes ConsistDreamer an instructionguided
scene editing framework that generates 3D consistently
edited images from 2D diffusion models Empirical evalua
tion shows that ConsistDreamer produces editing results of
significantly higher quality exhibiting sharper brighter ap
pearance with finegrained textures across various scenes