Hi, thanks for these nodes — really useful for 2.3 workflows.
The README currently frames this repo as built for LTX-2.3. With LTX-2.5 now natively supported in ComfyUI core (v0.32.0), I wanted to ask about compatibility for two nodes specifically:
LTX Tiled Sampler
This is a drop-in replacement for LTXVLatentUpsampler with spatial tiling. Since LTX-2.5 uses separate audio/video VAEs and the official 2.5 templates route through a different subgraph structure (with the new LTXV Dual CFG Guider, LTXV Spatio-Temporal Guidance, etc.), I'm wondering:
Does the tiling logic depend on anything specific to the 2.3 upscaler pipeline, or should it work as long as LTXVLatentUpsampler-compatible inputs are passed?
Has anyone tested the ×1.5/×2 auto-detection against 2.5 checkpoints/upscaler files?
LTX Latent Anchor Aware
This hooks into transformer_blocks[i].attn1. If the 2.5 transformer block architecture (dimensions, block count, or internal naming) changed from 2.3, this hook could silently fail or attach to the wrong tensor shape rather than throwing a clear error.
Has the block structure been verified against a 2.5 checkpoint?
Is there a known list of transformer_blocks attribute names/shapes for 2.5 vs 2.3 that would help confirm this?
Happy to test this myself on a 2.5 workflow and report back if that's useful — just wanted to check whether this is already a known unknown before I dig in, in case there's an obvious blocker (e.g. renamed attention modules) that would save me some debugging time.
Thanks!
Hi, thanks for these nodes — really useful for 2.3 workflows.
The README currently frames this repo as built for LTX-2.3. With LTX-2.5 now natively supported in ComfyUI core (v0.32.0), I wanted to ask about compatibility for two nodes specifically:
LTX Tiled Sampler
This is a drop-in replacement for LTXVLatentUpsampler with spatial tiling. Since LTX-2.5 uses separate audio/video VAEs and the official 2.5 templates route through a different subgraph structure (with the new LTXV Dual CFG Guider, LTXV Spatio-Temporal Guidance, etc.), I'm wondering:
Does the tiling logic depend on anything specific to the 2.3 upscaler pipeline, or should it work as long as LTXVLatentUpsampler-compatible inputs are passed?
Has anyone tested the ×1.5/×2 auto-detection against 2.5 checkpoints/upscaler files?
LTX Latent Anchor Aware
This hooks into transformer_blocks[i].attn1. If the 2.5 transformer block architecture (dimensions, block count, or internal naming) changed from 2.3, this hook could silently fail or attach to the wrong tensor shape rather than throwing a clear error.
Has the block structure been verified against a 2.5 checkpoint?
Is there a known list of transformer_blocks attribute names/shapes for 2.5 vs 2.3 that would help confirm this?
Happy to test this myself on a 2.5 workflow and report back if that's useful — just wanted to check whether this is already a known unknown before I dig in, in case there's an obvious blocker (e.g. renamed attention modules) that would save me some debugging time.
Thanks!