The purpose of this merge is surface rendering. Z-Image's spatial attention profile grafted onto Joy-LTX 2.5: straw with distinct strands, foliage that reads as leaves, physical ground litter, unretouched skin grain — identity, motion and sound untouched. Full method and calibrated expectations on the native repo; same-seed comparison stills attached.

Files (K-quant ladder, Q2_K → Q8_0)

file	size class	note
Q8_0	largest	quality reference
Q6_K / Q5_K_M	large	best quality-per-GB
Q4_K_M / Q4_K_S	mid	the RTX 30/40 sweet spot
Q3_K_M / Q2_K	small	low-VRAM, visible quality tax

On Ampere (RTX 30/40) these GGUFs run 4–8x faster than the emulated 4-bit comfy-native arms — check your ComfyUI load log's "Native ops:" line; if your card isn't listed as native for a quant op, you want this repo. On Blackwell the comfy-native int8/nvfp4 files (native repo) are the fast picks instead.

Drop-in for any Joy-LTX 2.5 v2 GGUF workflow — same loaders, same settings, same seeds as the stock v2 quants.

Same-seed comparisons (stock top row, graft bottom row - pause and look)

joeygambino/joyai-echo-ltx25-x-Z-Image-gguf

Joy-LTX 2.5 x Z-Image GGUF — surfaces, textures, fine details, foliage

Files (K-quant ladder, Q2_K → Q8_0)

Same-seed comparisons (stock top row, graft bottom row - pause and look)

joeygambino/
joyai-echo-ltx25-x-Z-Image-gguf