Initial support for Pixal3D - #38

Merged

pwilkin merged 22 commits into

pwilkin:mainfrom

vegax87:pixal3d

Sep 19, 2026

Commits on Aug 20, 2026

dit: thread a per-token projection input through the flow runner

claude
authored and
vegax87
committed
a541863
pixal3d: view-aligned projection conditioning and the NAF upsampler

claude
authored and
vegax87
committed
4486c5d
cli: select the model family with --model trellis|pixal3d

claude
authored and
vegax87
committed
176359e
convert + docs: Pixal3D model set and integration notes

claude
authored and
vegax87
committed
d7fdbf5
naf: refuse the unported pre-downsample branch; record verification

claude
authored and
vegax87
committed
7f79f09
tools: probe a GGUF's tensor table without downloading the weights

claude
authored and
vegax87
committed
0616731
tools: gguf_probe checks block structure and matmul dtypes

claude
authored and
vegax87
committed
47ae389
models: prefix the Pixal3D-specific GGUFs and share one model directory

claude
authored and
vegax87
committed
f371be6
convert: take model paths from the environment

claude
authored and
vegax87
committed
eb24b94
convert: let NAF_CKPT point straight at naf_release.pth

claude
authored and
vegax87
committed
0c6161c
docs: point at a pre-built Pixal3D GGUF set

claude
authored and
vegax87
committed
05f41cd
dit: name the tensor when a checkpoint's shapes do not fit the graph

claude
authored and
vegax87
committed
9b8dc63
tools: gguf_probe surfaces foreign-runtime metadata

claude
authored and
vegax87
committed
6f3c395
model: restore tensor shapes recorded by foreign-runtime converters

claude
authored and
vegax87
committed
1a05d81
pixal3d: report projection coverage and the two proj halves

claude
authored and
vegax87
committed
6b2ebf2
pixal3d: use Pixal3D's HR coord quantizer, not TRELLIS.2's

claude
authored and
vegax87
committed
a13146c
pixal3d: check the 1024 conditioning against the 512 one

claude
authored and
vegax87
committed
4d6676c
dit: skip FlashAttention when the key sequence is shorter than one tile

claude
authored and
vegax87
committed
b2ae9fa
pixal3d: expose --extend-pixel for subjects cropped by the frame

claude
authored and
vegax87
committed
a222b74
docs: correct what --extend-pixel is for

claude
authored and
vegax87
committed
78a2a36
server: accept the model family and the camera per request

claude
authored and
vegax87
committed
5dfb074
docs: record that proj mode depends on the FlashAttention V-range fix

claude
authored and
vegax87
committed
bb4a367

Navigation Menu

Uh oh!

Initial support for Pixal3D - #38

Initial support for Pixal3D#38
pwilkin merged 22 commits into
pwilkin:mainfrom
vegax87:pixal3d

Commits on Aug 20, 2026

dit: thread a per-token projection input through the flow runner

pixal3d: view-aligned projection conditioning and the NAF upsampler

cli: select the model family with --model trellis|pixal3d

convert + docs: Pixal3D model set and integration notes

naf: refuse the unported pre-downsample branch; record verification

tools: probe a GGUF's tensor table without downloading the weights

tools: gguf_probe checks block structure and matmul dtypes

models: prefix the Pixal3D-specific GGUFs and share one model directory

convert: take model paths from the environment

convert: let NAF_CKPT point straight at naf_release.pth

docs: point at a pre-built Pixal3D GGUF set

dit: name the tensor when a checkpoint's shapes do not fit the graph

tools: gguf_probe surfaces foreign-runtime metadata

model: restore tensor shapes recorded by foreign-runtime converters

pixal3d: report projection coverage and the two proj halves

pixal3d: use Pixal3D's HR coord quantizer, not TRELLIS.2's

pixal3d: check the 1024 conditioning against the 512 one

dit: skip FlashAttention when the key sequence is shorter than one tile

pixal3d: expose --extend-pixel for subjects cropped by the frame

docs: correct what --extend-pixel is for

server: accept the model family and the camera per request

docs: record that proj mode depends on the FlashAttention V-range fix