stream weights, dense objects, pixal3d and untextured runs

README.md

@@ -6,6 +6,7 @@ tags:
   - trellis
   - image-to-3d
   - low-vram
+  - pixal3d
 ---
 
 # over|yonder Espalier-2
@@ -20,9 +21,20 @@ An espalier is a fruit tree trained flat against a trellis, so a full-size tree
 ![Microsoft's published catapult (top) and Espalier-2's at 1536 on a 16 GB card (bottom)](catapult-comparison.jpg)
 
 Microsoft lists a 24 GB GPU as the minimum, and on a 16 GB card its 1536 pipeline runs out of memory in the shape
-decoder and again in the GLB export. Espalier-2 tiles the decoders' two highest-resolution stages, runs the export's
-remesh in chunks, and keeps data the export doesn't need off the GPU. On an RTX 5070 Ti, generating the catapult above
-peaked at 11.9 GiB and its export at 13.3 GiB, measured from the NVIDIA driver.
+decoder and again in the GLB export. Espalier-2 tiles the decoders' highest-resolution stages, runs the export's remesh
+in chunks, and frees memory the remesh and simplifier no longer need. On an RTX 5070 Ti, generating the catapult above
+peaked at 11.9 GiB and its export at 8.3 GiB, measured from the NVIDIA driver.
+
+Dense objects also stay at 1536. Microsoft's cascade lowers the resolution of objects above 49,152 tokens, and
+Espalier-2's `--max-tokens` option raises that limit: a water pot of 80,816 tokens generated at 1536 with a 13.0 GiB
+peak and exported at 11.5 GiB.
+
+Each network's weights are read from its safetensors file to the GPU when its stage runs, so the process holds about
+6 GiB of system memory instead of 20 GiB, with bit-identical output.
+
+Espalier-2 also runs Tencent's [Pixal3D](https://huggingface.co/TencentARC/Pixal3D), which is built on TRELLIS.2, at
+1536 on the same card, from one image or from several posed views. With `--no-texture`, either model skips the texture
+and exports the untextured geometry, which is identical to a textured run's.
 
 ![The same patch of both catapults from a 3072-pixel render](catapult-closeup.jpg)