JoyAI-Echo x LTX-2.5 (echoVid) - comfy-native (int8 / w4a8 / w4a4 / nvfp4 / mixed)
Try it in the browser: ZeroGPU demo Space - these exact files, the two-pass ladder, no install.
LTX-2.5's engine with JoyAI-Echo's performance. LTX-2.5 renders picture and sound in
one pass, at any length, in one generation. JoyAI-Echo (a fine-tune of LTX-2.3) has the
better
actor: natural lip-sync, expressive faces, a voice that stays put. The two
transformers are shape-identical, so JoyAI-Echo's video attention/feed-forward delta was
transplanted onto the official
LTX-2.5 dev transformer, and the official LTX-2.5 distilled LoRA (
ltx-2.5-22b-distilled-lora-450) is baked in at 0.5 - so these are few-step files with the same speed, VRAM and nodes as LTX-2.5 distilled. Nothing was retrained. (
v2: the first build put the delta on the distilled transformer and came out over-saturated with hard contrast; those files are gone. The plain dev merges, for people who want to apply their own distill LoRA at their own strength, are here:
https://huggingface.co/joeygambino/joyai-echo-ltx25-echoVid-dev.)
What you get over stock LTX-2.5 distilled is the acting JoyAI-Echo was trained for - lip-sync,
expression, a voice that stays put - at the same speed, VRAM and nodes.
Workflow + nodes: https://github.com/jlucasmcrell/ComfyUI-JoyLTX25 (the
Joy-LTX 2.5
canvases: one-prompt take with a VRAM planner, and multishot with AV-extend joins; the release zip bundles the writer).
GGUF files (Q3_K_M .. Q8_0):
https://huggingface.co/joeygambino/joyai-echo-ltx25-echoVid-gguf
All models: https://huggingface.co/joeygambino
Try it live: https://huggingface.co/spaces/joeygambino/joy-ltx-25 (one take, ZeroGPU)
Civitai: Joy-LTX 2.5 (models being uploaded now).
What it looks like
Rendered with the files on this page (070T30, distilled LoRA baked at 0.5), the
ComfyUI-JoyLTX25 canvases, 8 steps at cfg 1. Sound is generated with the picture, in the same pass - turn it on.
Three shots joined into one take
Multishot, 3 x 8 s at 1280x736, AV-extend joins - the speech and the room carry across both joins with no reference photo attached.
Beach, hard sun
10 s, single generation, picture and sound together.
Wet neon street
10 s, single generation. Reflections and rain with a voice over them.
Snow, flat overcast
10 s, single generation. The grade holds in high key - the failure mode of the first build.
Two doses
| dose | what it is | pick it when |
|---|
| 070T30 (default) | 0.7 x Echo delta on video attention/FF, 0.3 x on the modulation tables, distill LoRA 0.5 | the default - cleaner skin, natural grade |
| 100T50 (strong) | 1.0 x / 0.5 x, distill LoRA 0.5 | loud, comic, animated performances - the livelier read, a touch hotter on contrast |
Both were reviewed blind on 20+ paired renders: scores tie; 070T30 reads a touch less
rubbery on still faces, 100T50 lands laughter and big expressions better. Start with 070T30.
Which file (stock ComfyUI 0.32+, no custom loader - the fast family on RTX 50)
These use ComfyUI's own quantisation (comfy_quant + comfy-kitchen kernels), the same
machinery as Lightricks' official int8-convrot build. Load them with the plain Load
Diffusion Model node. Sizes are decimal GB. Timings: 960x544, 8 s, two-pass x2 to 1920x1088.
| file | GB | fits | RTX 5090 | RTX 3090 |
|---|
LTX25dist-echoVid-<dose>-v2-DiT-comfy-w4a4.safetensors | 11.2 | 12 GB (tight) / 16 GB | 87 s | 3121 s (avoid on Ampere) |
LTX25dist-echoVid-<dose>-v2-DiT-comfy-w4a8.safetensors | 12.5 | 16 GB | ~90 s | ~580 s |
LTX25dist-echoVid-<dose>-v2-DiT-comfy-nvfp4.safetensors | 12.5 | 16 GB (RTX 50 only) | ~100 s | n/a |
LTX25dist-echoVid-<dose>-v2-DiT-comfy-mix4x8-13.8GB.safetensors | 13.8 | 16 GB | 110 s | 1685 s |
LTX25dist-echoVid-<dose>-v2-DiT-comfy-mix4x8-17.0GB.safetensors | 17.0 | 24 GB | 111 s | 3093 s (offloads) |
LTX25dist-echoVid-<dose>-v2-DiT-comfy-int8.safetensors | 21.5 | 32 GB (24 GB tight) | 120 s (32 GB default) | - |
Rule of thumb: RTX 50 -> this repo. RTX 30/40 -> the GGUF repo (Q5_K_M / Q6_K are 4-8x
faster there than any 4-bit comfy-native arm). --enable-triton-backend on the ComfyUI
launch line roughly halves w4a8/int8 step time where triton is installed.
fp8 and the bf16 master
Two more cuts, straight from the v2 master (same bake: dev + Echo delta + distill LoRA 0.5):
| file | GB | note |
|---|
LTX25dist-echoVid-<dose>-v2-DiT-comfy-fp8.safetensors | 21.5 | comfy fp8_e4m3fn scaled; stock Load Diffusion Model |
LTX25dist-echoVid-<dose>-v2-DiT-bf16.safetensors | 42.0 | the master; needs a card that streams 42 GB (or offload); the file to quantise from |
Install (ComfyUI)
- ComfyUI 0.32 or newer (the comfy-kitchen kernels ship with it).
- Put the
.safetensors in models/diffusion_models/.
- From Lightricks/LTX-2.5:
vae/ltx-2.5-video-vae-bf16.safetensors
and vae/ltx-2.5-audio-vae-bf16.safetensors -> models/vae/;
latent_upscale_models/ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors -> models/latent_upscale_models/;
text encoder text_encoders/gemma4-12b-with-proj-ltx-2.5-comfy-int8-convrot.safetensors -> models/text_encoders/
(16 GB cards: the 10.6 GB gemma4-12b-ltx25-comfy-w4a8.safetensors from
LTX-2.5-Quantized).
- Load the workflow from the node pack above (or any LTX-2.5 workflow: pick this file in the
stock Load Diffusion Model loader). Distilled schedule: 8 steps pass 1, 3 steps pass 2,
euler_ancestral, CFG 1.
Credits
JoyAI-Echo by JD (
jdopensource/JoyAI-Echo);
LTX-2.5 by Lightricks. Merge, quantisation and workflows by joeygambino. Licensed under the
LTX-2.x Community License (inherited from both parents).
