joyai-echo-ltx25-echoVid-comfy-native

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

<video src="https://huggingface.co/joeygambino/joyai-echo-ltx25-echoVid-comfy-native/resolve/main/demo/demo_seamless_3shot_23s.mp4" controls width="640"></video>

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

<video src="https://huggingface.co/joeygambino/joyai-echo-ltx25-echoVid-comfy-native/resolve/main/demo/demo_beach_10s.mp4" controls width="640"></video>

10 s, single generation, picture and sound together.

Wet neon street

<video src="https://huggingface.co/joeygambino/joyai-echo-ltx25-echoVid-comfy-native/resolve/main/demo/demo_neon_10s.mp4" controls width="640"></video>

10 s, single generation. Reflections and rain with a voice over them.

Snow, flat overcast

<video src="https://huggingface.co/joeygambino/joyai-echo-ltx25-echoVid-comfy-native/resolve/main/demo/demo_snow_10s.mp4" controls width="640"></video>

10 s, single generation. The grade holds in high key - the failure mode of the first build.

Two doses

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.

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):

Install (ComfyUI)

1. ComfyUI 0.32 or newer (the comfy-kitchen kernels ship with it).
2. Put the .safetensors in models/diffusion_models/.
3. 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).
4. 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).