Repository navigation
Gallery #19
Description
Activity
- pinned this issue
on Jun 17, 2023 "Michelangelo style statue of dog reading news on a cellphone"
performance: 15 minutes for geometry modeling and 13 minutes for appearance modeling on 8 3090 GPUs
resolution: 256dog_appearance_stratege0.mp4
dog_geometry.mp4
Reacted by Rui Chen, Yongwei Chen and xmu-xiaoma666"A highly detailed sandcastle"
performance: 15 minutes for geometry modeling and 13 minutes for appearance modeling on 8 3090 GPUs
resolution: 256sandcastle_geometry.mp4
sandcastle_appearacne_strategy0.mp4
sandcastle_appearance_strategy1.mp4
sandcastle_appearance_strategy2.mp4
Reacted by Rui Chen, Yongwei Chen and xmu-xiaoma666"a baby bunny sitting on top of a stack of pancakes".
performance: 15 minutes for geometry modeling and 13 minutes for appearance modeling on 8 3090 GPUs
resolution: 256See Google Drive to visualize the training process.
baby_bunny_geometry.mp4
bunny_appearance_strategy0.mp4
Reacted by Rui Chen and Yongwei ChenHi, could you also provide model results for downloading (obj file, or other format)?
Hi, @akturtle ! No problem. But there may only be geometric parts, as UV padding for appearance requires some additional time using blender.
Thanks for your great work. I cannot reproduce the same results for I have only 4 GPUs and I met the Jesus multi-head problems. Could you release the generated ckpts for the gallery such as the trained weight of mlptexture?
@colorful-liyu Which config file did you use and what text did you use? I have been preparing for my final exam recently, so some requirements will only be realized after I finish the exam.
@colorful-liyu Which config file did you use and what text did you use? I have been preparing for my final exam recently, so some requirements will only be realized after I finish the exam.
Specifically, the geometry generation results of bama_head and oldman always have multi-face or three ears, while the generated elephant has separate ivory. And the dog reading newspapers may have more than four legs. The appearance generation looks nice. So any release about .obj files especially ckpts will be helpful.
Hi @ARuiChen,
Great work! Could you please help answer a quick question? What's the average speedup you see in 8GPU vs 1GPU? Thank you so much!
"A high heel made of gold".
performance: 7 minutes for geometry modeling and 5 minutes for appearance modeling on 8 A100 GPUs
DMTet resolution: 256
zero_shot generationheel_appearance.mp4
Reacted by AI enthusiastReacted by Rui Chen"A DSLR photo of hamburger".
performance: 7 minutes for geometry modeling and 12 minutes for appearance modeling on 1 A100 GPU
DMTet resolution: 256
zero_shot generationhamburger_appearance.mp4
Reacted by AI enthusiast and Sang Min KimReacted by Rui Chen@colorful-liyu When you meet multiple faces, removing the annotation here and rotating the initialized shape with some angles can greatly alleviate janus-problem. But it is important to note that adding negative text in some cases may cause poor deformation ability.
"a full body photo of a girl wearing a mecha"
performance: 11 minutes for geometry modeling and 8 minutes for appearance modeling on 8 A100 GPUs
DMTet resolution: 256
user-guided generation
See Google Drive to visualize the training process.girl_geometry.mp4
girl_appearance.mp4
cyber_girl_appeappearance.mp4
Reacted by Rui Chen18 remaining items
@VLadImirluren . Of course, except for some that I don't want to release yet, I will release most of the corresponding configurations in the future. But I am currently busy with something and I will finish it tomorrow evening.
sometimes, I can't reach the same as yours.I am confused that is the problem of different GPU? Can you share the GPU you use with us?(I see most of them are A100 * 8, how can I reach the same as yours on V100 , 3090 or 4090?)
The GPU type only affects the generation speed and does not have a significant impact on the results. The most significant impact may be on the number of GPUs, as my algorithm was debugged under 8 GPUs， which is related to the sampling perspective in the dataset.py.
In addition, the random numbers generated in different machines may vary using the same seed, which may also be one of the influencing factors.
You can publish a separate issue about your generated results and corresponding configurations, and I will take some time to make suggestions on this.
Lastly, please ensure that your code is up to date.The GPU type only affects the generation speed and does not have a significant impact on the results. The most significant impact may be on the number of GPUs, as my algorithm was debugged under 8 GPUs， which is related to the sampling perspective in the dataset.py. In addition, the random numbers generated in different machines may vary using the same seed, which may also be one of the influencing factors. You can publish a separate issue about your generated results and corresponding configurations, and I will take some time to make suggestions on this. Lastly, please ensure that your code is up to date.
thanks
@ARuiChen I went back and tested for a long time, but I still couldn't reproduce the appearance of the cottage. By adjusting the strategy, I made the geometry smoother, but there is still a lot of noise in the appearance. Could you please share the appearance settings of the cottage? Thanks
Sure, wait a few minutes for me.
Sure, wait a few minutes for me.
Could you please show the geometry and appearance of "a model of a house in Tudor style" if it is convenient? . . thanks. . . . .
Done. house_geometry.json and house_appearance.json. Please pull the latest code. Push another issue if you have any other questions. This is where the Gallery is located ^_^.
Done. house_geometry.json and house_appearance.json. Please pull the latest code. Push another issue if you have any other questions. This is where the Gallery is located ^_^.
thanks. I will
“A man riding a bike”
Performance: 9 minutes for geometry modeling on 8 A100 GPUs
DMTet resolution: 256bike_geometry.mp4
squirrel_geometry1.json and squirrel_appearance1_strategy2.json
Adding negative prompts can improve the visual quality in appearance modeling!
text: “A squirrel”
negative text: "shadow, oversaturated color, strange color"
Performance: 7 minutes for geometry modeling and 8 minutes for appearance modeling on 8 A100 GPUs
DMTet resolution: 256If the negative prompt is not added, in some cases, you can see many shadows and strange colors:
w/ the negative prompt ( just used in appearance modeling):
squirrel_geometry1.mp4
squrrel_appearance_strategy2.mp4
text: “michelangelo style statue of an astronaut carrying space jet backpacks and wearing spacesuits”
negative text: "shadow, oversaturated color, strange color"
Performance: 8 minutes for geometry modeling and 8 minutes for appearance modeling on 8 A100 GPUs
DMTet resolution: 256michelangelo_astronaut_geometry.mp4
michelangelo_astronaut_appearance_strategy1.mp4
michelangelo_astronaut_appearance_strategy2.mp4
text: “A standing bulldog wearing a black pirate hat”
Performance: 8 minutes for geometry modeling and 8 minutes for appearance modeling on 8 A100 GPUs
DMTet resolution: 256bulldog_geometry.mp4
bulldog_appearance_strategy0.mp4
bulldog_appearance_strategy2.mp4
text: “military Mech, long barrel”
Performance: 8 minutes for geometry modeling and 8 minutes for appearance modeling on 8 A100 GPUs
DMTet resolution: 256military_geometry.mp4
military_appearance_strategy2.mp4
