The purpose of this merge is surface rendering. Z-Image's spatial attention profile is grafted onto the Joy-LTX 2.5 engine so that straw holds individual strands at mid-distance, foliage reads as leaves instead of green mass, ground litter stays physical, and skin keeps its unretouched grain — while identity, motion, sound and the engine's behavior stay exactly what Joy-LTX 2.5 already was.

This is the same marriage already shipped for MiniMax-H3 (MiniMax-H3-x-Z-Image-native), applied to the LTX-2.5 line: Z-Image (Lumina2, 6B, exceptional texture rendering) donates a per-block statistics profile — row norms of its fused-QKV q-slice, mean-normalized, depth-interpolated to 48 blocks — which rescales this DiT's own q_norm weights by small per-block factors (clip [0.85, 1.18], dose 0.5). No retraining, no new knowledge, no architecture change, no Z-Image weights at inference.

Same-seed A/Bs against stock Joy-LTX 2.5 v2 show a consistent, measurable surface win: +8–23% fine-detail energy on matched frames, clearest on hay, foliage, gravel and fabric at mid-distance. It is a paused-frame visible improvement — side-by-side stills show it plainly; in motion it reads as "richer" rather than dramatic. Blind reviewer panels score both arms close, which is exactly the intent: nothing about the render changes except how well surfaces resolve. Comparison stills are attached from the same-seed pairs.

RTX 30/40 (Ampere): use the GGUF repo instead — K-quants run 4–8x faster than any emulated 4-bit arm there.

Drop-in for any Joy-LTX 2.5 v2 workflow — same loaders, same settings, same seeds.