-
Add workflow setup notes (files, node deps, sampler rationale)
-
Add the workflow.json the README promised (h3_fasth3_T1.safetensors, 4 steps, euler)
-
METHODOLOGY: measured max clip length per config (dense 310f/124f; VSA 68f/OOM)
-
METHODOLOGY: rotation A/B as prose explanation, drop the table
-
METHODOLOGY: replace hand-wave with measured same-weights rotation A/B table + honest ~15% cost + parity-vs-community speed; soften un-rotated=mushy claim
-
loader: build fp4 blocks directly (no ~72GB bf16 load transient); --gpu-only loads in 4s @10.57GB peak, DiT resident 12.777GB zero offload; output bit-identical (SSIM 1.0). Device-bug fix retained.
-
H3RotNVFP4Loader node + methodology (device-bug fixed)
-
initial commit
