Nicolas Barradeau
@nicoptere.bsky.social
creative coder https://www.barradeau.com
looks good :)
not sure what the glyphs are but this type of layout exploration always reminds me of a 2014 article by Flipboard fr-fr.about.flipboard.com/2014/03/24/l... where they explain how they create and choose the layouts ( cdn.flipboard.com/wp-content/u... cdn.flipboard.com/wp-content/u... )
not sure what the glyphs are but this type of layout exploration always reminds me of a 2014 article by Flipboard fr-fr.about.flipboard.com/2014/03/24/l... where they explain how they create and choose the layouts ( cdn.flipboard.com/wp-content/u... cdn.flipboard.com/wp-content/u... )
June 26, 2025 at 7:30 AM
looks good :)
not sure what the glyphs are but this type of layout exploration always reminds me of a 2014 article by Flipboard fr-fr.about.flipboard.com/2014/03/24/l... where they explain how they create and choose the layouts ( cdn.flipboard.com/wp-content/u... cdn.flipboard.com/wp-content/u... )
not sure what the glyphs are but this type of layout exploration always reminds me of a 2014 article by Flipboard fr-fr.about.flipboard.com/2014/03/24/l... where they explain how they create and choose the layouts ( cdn.flipboard.com/wp-content/u... cdn.flipboard.com/wp-content/u... )
another image to 3D model, excellent NV Synth, more faithful to the source image than TRELLIS, max 1536^3 resolution and it creates watertight meshes (=3D print ready)! lizhihao6.github.io/Sparc3D/ hf demo: huggingface.co/spaces/ilcve...
not sure about the GPU requirements, probably high...
not sure about the GPU requirements, probably high...
June 13, 2025 at 8:04 AM
another image to 3D model, excellent NV Synth, more faithful to the source image than TRELLIS, max 1536^3 resolution and it creates watertight meshes (=3D print ready)! lizhihao6.github.io/Sparc3D/ hf demo: huggingface.co/spaces/ilcve...
not sure about the GPU requirements, probably high...
not sure about the GPU requirements, probably high...
3D model generation often produces monolithic models with no semantics. here are some efforts in trying to make sense of the a 3D model to make them more usable (untested)
PartField
github.com/nv-tlabs/Par...
SAMPart3D
github.com/Pointcept/SA...
Primitive Anything
github.com/PrimitiveAny...
PartField
github.com/nv-tlabs/Par...
SAMPart3D
github.com/Pointcept/SA...
Primitive Anything
github.com/PrimitiveAny...
May 8, 2025 at 11:42 AM
3D model generation often produces monolithic models with no semantics. here are some efforts in trying to make sense of the a 3D model to make them more usable (untested)
PartField
github.com/nv-tlabs/Par...
SAMPart3D
github.com/Pointcept/SA...
Primitive Anything
github.com/PrimitiveAny...
PartField
github.com/nv-tlabs/Par...
SAMPart3D
github.com/Pointcept/SA...
Primitive Anything
github.com/PrimitiveAny...
the "Picture of Dorian Gray" rendered using Gemini-2's new inlined image generator (to test the consistency)
March 18, 2025 at 2:51 PM
the "Picture of Dorian Gray" rendered using Gemini-2's new inlined image generator (to test the consistency)
saw some pretty bad AI generated images on a "naturopathy" website today (it was recommended but muzli) and I think I'll never get used to people using generated images "as is", put them online and call it a day.
March 16, 2025 at 2:44 PM
saw some pretty bad AI generated images on a "naturopathy" website today (it was recommended but muzli) and I think I'll never get used to people using generated images "as is", put them online and call it a day.
there are tools that can convert a GS model to 3D mesh like: github.com/GAP-LAB-CUHK...
probably worth trying with the exported 3DGS to see if it gives better results.
because of the SLat and its voxel grid structure, some noise may be introduced though. visually, it gives evenly spaced grey splats
probably worth trying with the exported 3DGS to see if it gives better results.
because of the SLat and its voxel grid structure, some noise may be introduced though. visually, it gives evenly spaced grey splats
January 19, 2025 at 10:05 AM
there are tools that can convert a GS model to 3D mesh like: github.com/GAP-LAB-CUHK...
probably worth trying with the exported 3DGS to see if it gives better results.
because of the SLat and its voxel grid structure, some noise may be introduced though. visually, it gives evenly spaced grey splats
probably worth trying with the exported 3DGS to see if it gives better results.
because of the SLat and its voxel grid structure, some noise may be introduced though. visually, it gives evenly spaced grey splats
skeletons are tricky as fuck, wasted a day trying out various methods. most libs are not maintained or don't produce useful results, or only work on watertight meshes .... I ended up with a handmade voxel approach & have to simplify & vectorize the graph which is exactly what I wanted to avoid :/
January 17, 2025 at 10:30 PM
skeletons are tricky as fuck, wasted a day trying out various methods. most libs are not maintained or don't produce useful results, or only work on watertight meshes .... I ended up with a handmade voxel approach & have to simplify & vectorize the graph which is exactly what I wanted to avoid :/
kinda getting somewhere ; at least now the convex hulls enclose somewhat meaningful parts of the mesh... next step would be to merge patches (connected components) based on ... distance to hull? I can think of so many edge cases already... 🙄
January 15, 2025 at 9:31 PM
kinda getting somewhere ; at least now the convex hulls enclose somewhat meaningful parts of the mesh... next step would be to merge patches (connected components) based on ... distance to hull? I can think of so many edge cases already... 🙄
unsurprisingly Mv-DUSt3R ( github.com/facebookrese...) does not like 'turntable' footage and prefers continuous sequences. it can't replace - say - TRELLIS to convert few shots of an object into a 3D model.
it's fast though, not sure how to prune discretize the result ( nice debug views too ^^)
it's fast though, not sure how to prune discretize the result ( nice debug views too ^^)
January 15, 2025 at 12:11 PM
unsurprisingly Mv-DUSt3R ( github.com/facebookrese...) does not like 'turntable' footage and prefers continuous sequences. it can't replace - say - TRELLIS to convert few shots of an object into a 3D model.
it's fast though, not sure how to prune discretize the result ( nice debug views too ^^)
it's fast though, not sure how to prune discretize the result ( nice debug views too ^^)
TRELLIS works especially well on mechanical objects. in this case it made the mesh roughly symmetrical and fixed a missing leg.
all the components are disjoint 'patches' scattered all over the mesh, disregarding the topology. I wonder how useful this can be for 3D artists, maybe as static assets?
all the components are disjoint 'patches' scattered all over the mesh, disregarding the topology. I wonder how useful this can be for 3D artists, maybe as static assets?
January 14, 2025 at 8:48 PM
TRELLIS works especially well on mechanical objects. in this case it made the mesh roughly symmetrical and fixed a missing leg.
all the components are disjoint 'patches' scattered all over the mesh, disregarding the topology. I wonder how useful this can be for 3D artists, maybe as static assets?
all the components are disjoint 'patches' scattered all over the mesh, disregarding the topology. I wonder how useful this can be for 3D artists, maybe as static assets?
to learn how to make an Blender add-on, I made a TRELLIS-Blender bridge.
Blender is a powerful tool and not so horrible once you give up all hope and go with the (stack over)flow. wish I could explore further :)
the add-on available here github.com/nicoptere/tr...
Blender is a powerful tool and not so horrible once you give up all hope and go with the (stack over)flow. wish I could explore further :)
the add-on available here github.com/nicoptere/tr...
January 10, 2025 at 10:03 PM
to learn how to make an Blender add-on, I made a TRELLIS-Blender bridge.
Blender is a powerful tool and not so horrible once you give up all hope and go with the (stack over)flow. wish I could explore further :)
the add-on available here github.com/nicoptere/tr...
Blender is a powerful tool and not so horrible once you give up all hope and go with the (stack over)flow. wish I could explore further :)
the add-on available here github.com/nicoptere/tr...
forgot the material 'ply_preview' is required too...
January 10, 2025 at 3:16 PM
forgot the material 'ply_preview' is required too...
I managed to display a colored point cloud in Blender! it took longer than anticipated but it can be handy so I'll leave screenshots here.
the import script is here pastebin.com/Dz2VusYH
the PLY I used is a Gaussian Splat so it looks like shit and the R, G & B attributes are f_dc_0, f_dc_1 & f_dc_2
the import script is here pastebin.com/Dz2VusYH
the PLY I used is a Gaussian Splat so it looks like shit and the R, G & B attributes are f_dc_0, f_dc_1 & f_dc_2
January 9, 2025 at 3:20 PM
I managed to display a colored point cloud in Blender! it took longer than anticipated but it can be handy so I'll leave screenshots here.
the import script is here pastebin.com/Dz2VusYH
the PLY I used is a Gaussian Splat so it looks like shit and the R, G & B attributes are f_dc_0, f_dc_1 & f_dc_2
the import script is here pastebin.com/Dz2VusYH
the PLY I used is a Gaussian Splat so it looks like shit and the R, G & B attributes are f_dc_0, f_dc_1 & f_dc_2
immediately replaced by github.com/IGL-HKUST/Di... (at least they announced a release date this time :) )
not convinced by mesh to video & motion transfer but the camera control & object manipulation look good
+ it has "shader" in the name....
not convinced by mesh to video & motion transfer but the camera control & object manipulation look good
+ it has "shader" in the name....
January 9, 2025 at 9:14 AM
immediately replaced by github.com/IGL-HKUST/Di... (at least they announced a release date this time :) )
not convinced by mesh to video & motion transfer but the camera control & object manipulation look good
+ it has "shader" in the name....
not convinced by mesh to video & motion transfer but the camera control & object manipulation look good
+ it has "shader" in the name....
poor thing... 😅
I suspect the headpiece is interpreted as a foreground plane and breaks the MV synth. pretty decent modeling
( image from civit.ai, forgot to bookmark :/ )
I suspect the headpiece is interpreted as a foreground plane and breaks the MV synth. pretty decent modeling
( image from civit.ai, forgot to bookmark :/ )
December 19, 2024 at 11:15 PM
poor thing... 😅
I suspect the headpiece is interpreted as a foreground plane and breaks the MV synth. pretty decent modeling
( image from civit.ai, forgot to bookmark :/ )
I suspect the headpiece is interpreted as a foreground plane and breaks the MV synth. pretty decent modeling
( image from civit.ai, forgot to bookmark :/ )
saw a ClipDrop mention in the wild, put a smile on my face
here => github.com/pengHTYX/Era...
cc @damien-henry.bsky.social & @cyrildiagne.bsky.social
here => github.com/pengHTYX/Era...
cc @damien-henry.bsky.social & @cyrildiagne.bsky.social
December 11, 2024 at 2:50 PM
saw a ClipDrop mention in the wild, put a smile on my face
here => github.com/pengHTYX/Era...
cc @damien-henry.bsky.social & @cyrildiagne.bsky.social
here => github.com/pengHTYX/Era...
cc @damien-henry.bsky.social & @cyrildiagne.bsky.social
it's outputs a properly uv mapped mesh with a texture :)
and pretty optimized too: 40K vertices / 50K tris for this model is very decent.
the source image needs to be - somehow - "realistic" though ; I tried with a cell-shaded character and it produced a quad...
and pretty optimized too: 40K vertices / 50K tris for this model is very decent.
the source image needs to be - somehow - "realistic" though ; I tried with a cell-shaded character and it produced a quad...
December 8, 2024 at 7:26 PM
it's outputs a properly uv mapped mesh with a texture :)
and pretty optimized too: 40K vertices / 50K tris for this model is very decent.
the source image needs to be - somehow - "realistic" though ; I tried with a cell-shaded character and it produced a quad...
and pretty optimized too: 40K vertices / 50K tris for this model is very decent.
the source image needs to be - somehow - "realistic" though ; I tried with a cell-shaded character and it produced a quad...
this is the source image: civitai.com/images/43976...
there's a background removal step then it produces the video from which it can build the textured 3D model.
there's a background removal step then it produces the video from which it can build the textured 3D model.
December 8, 2024 at 2:44 PM
this is the source image: civitai.com/images/43976...
there's a background removal step then it produces the video from which it can build the textured 3D model.
there's a background removal step then it produces the video from which it can build the textured 3D model.
TRELLIS slaps reeeeally hard! :O
still many inconsiotencies and some 'unfinished business' in places but overall it does a great job (source image in next post)
test here huggingface.co/spaces/Jeffr...
repo here: github.com/microsoft/TR...
still many inconsiotencies and some 'unfinished business' in places but overall it does a great job (source image in next post)
test here huggingface.co/spaces/Jeffr...
repo here: github.com/microsoft/TR...
December 8, 2024 at 2:44 PM
TRELLIS slaps reeeeally hard! :O
still many inconsiotencies and some 'unfinished business' in places but overall it does a great job (source image in next post)
test here huggingface.co/spaces/Jeffr...
repo here: github.com/microsoft/TR...
still many inconsiotencies and some 'unfinished business' in places but overall it does a great job (source image in next post)
test here huggingface.co/spaces/Jeffr...
repo here: github.com/microsoft/TR...
motion-prompting.github.io looks good! the "failure" video is hilarious too :)
via @dreamingtulpa.bsky.social => bsky.app/profile/did:...
via @dreamingtulpa.bsky.social => bsky.app/profile/did:...
December 6, 2024 at 3:51 PM
motion-prompting.github.io looks good! the "failure" video is hilarious too :)
via @dreamingtulpa.bsky.social => bsky.app/profile/did:...
via @dreamingtulpa.bsky.social => bsky.app/profile/did:...
spent the last two days trying out WebGPU and though I knew WGSL was more picky than GLSL, didn't expect it to be _-THAT-_ picky 😱
I painstakingly managed to get this glorious piece of SDF running, not sure I'll venture much further into Vanilla WebGPU...
I painstakingly managed to get this glorious piece of SDF running, not sure I'll venture much further into Vanilla WebGPU...
December 5, 2024 at 11:37 AM
spent the last two days trying out WebGPU and though I knew WGSL was more picky than GLSL, didn't expect it to be _-THAT-_ picky 😱
I painstakingly managed to get this glorious piece of SDF running, not sure I'll venture much further into Vanilla WebGPU...
I painstakingly managed to get this glorious piece of SDF running, not sure I'll venture much further into Vanilla WebGPU...
RIley Walz (@riley.bsky.social) made a candid and somewhat moving thing. in the early days of iPhone (2009-2012), users could easily push videos to youtube and most of the time, the videos kept their default name (IMG_####).
RIley collected them and now you can watch them here: walzr.com/IMG_0001
RIley collected them and now you can watch them here: walzr.com/IMG_0001
November 19, 2024 at 11:29 AM
RIley Walz (@riley.bsky.social) made a candid and somewhat moving thing. in the early days of iPhone (2009-2012), users could easily push videos to youtube and most of the time, the videos kept their default name (IMG_####).
RIley collected them and now you can watch them here: walzr.com/IMG_0001
RIley collected them and now you can watch them here: walzr.com/IMG_0001
Microsoft captured and rendered Saint peter's Basilica as Gaussian Splatting model.
it's pretty massive (they used ~400K images) and rather beautiful.
check it out here:
virtualbasilicaprod.azurewebsites.net/en/basilica-...
source: radiancefields.com/radiance-fie...
it's pretty massive (they used ~400K images) and rather beautiful.
check it out here:
virtualbasilicaprod.azurewebsites.net/en/basilica-...
source: radiancefields.com/radiance-fie...
November 15, 2024 at 11:21 AM
Microsoft captured and rendered Saint peter's Basilica as Gaussian Splatting model.
it's pretty massive (they used ~400K images) and rather beautiful.
check it out here:
virtualbasilicaprod.azurewebsites.net/en/basilica-...
source: radiancefields.com/radiance-fie...
it's pretty massive (they used ~400K images) and rather beautiful.
check it out here:
virtualbasilicaprod.azurewebsites.net/en/basilica-...
source: radiancefields.com/radiance-fie...
a demo by @fffiloni.bsky.social is available on huggingface @huggingface.bsky.social => huggingface.co/spaces/fffil...
can't wait to try and splat some!
can't wait to try and splat some!
November 12, 2024 at 4:16 PM
a demo by @fffiloni.bsky.social is available on huggingface @huggingface.bsky.social => huggingface.co/spaces/fffil...
can't wait to try and splat some!
can't wait to try and splat some!
top picks from @arnicas.bsky.social TITAA
- Robert Hodgin's article on maps👌 roberthodgin.com/project/goog...
- Victoria Turk's biases in text to images models restofworld.org/2023/ai-imag...
- foolig multimodal LLMs (the "watermark" attack vector is ace, see pic) simonwillison.net/2023/Oct/14/...
- Robert Hodgin's article on maps👌 roberthodgin.com/project/goog...
- Victoria Turk's biases in text to images models restofworld.org/2023/ai-imag...
- foolig multimodal LLMs (the "watermark" attack vector is ace, see pic) simonwillison.net/2023/Oct/14/...
October 16, 2023 at 8:23 AM
top picks from @arnicas.bsky.social TITAA
- Robert Hodgin's article on maps👌 roberthodgin.com/project/goog...
- Victoria Turk's biases in text to images models restofworld.org/2023/ai-imag...
- foolig multimodal LLMs (the "watermark" attack vector is ace, see pic) simonwillison.net/2023/Oct/14/...
- Robert Hodgin's article on maps👌 roberthodgin.com/project/goog...
- Victoria Turk's biases in text to images models restofworld.org/2023/ai-imag...
- foolig multimodal LLMs (the "watermark" attack vector is ace, see pic) simonwillison.net/2023/Oct/14/...