v1.6.1

Added

• EXRColorSpace.ACES2065_1 (aces2065-1 / ACES2065_1) and Primaries.AP0 add SMPTE ST 2065-1 (AP0 linear, D60 whitepoint) input and output support on --hdr, --exr-colorspace, and --input-colorspace. ACES2065-1 EXR outputs are written to <stem>_aces2065-1_exr/ as SMPTE ST 2065-4 compliant uncompressed half-float containers (compression="none", acesImageContainerFlag=1, adoptedNeutral).

Changed

• DubItPipeline.__call__ and TI2VidTwoStagesHQPipeline.__call__ no longer enter torch.inference_mode() themselves, like the other single-GPU pipelines, so they can run where gradients are recorded. The ltx_pipelines.dubit and ltx_pipelines.ti2vid_two_stages_hq CLIs still run under inference_mode; Python callers should wrap the call in with torch.inference_mode(): (see Calling pipelines from Python).
• to_hdr_linear clamps sub-black decompressed ACEScct values at 0.0 in AP1 before converting to out_primaries (so negative AP1 undershoot below the ACEScct toe intercept no longer mixes across channels when converting to Rec.709 for srgb_linear EXRs and HLG masters), and preserves negative out-of-AP1-gamut lobes when out_primaries=Primaries.AP0.
• The ltx_kernels_inductor compile backend also fuses bias-free linear projections followed by tanh GELU (new LinearGELUNoBiasPass), such as the LTX-2.5 video feed-forward. ltx_kernels::linear_gelu accepts bias=None.