Joy-LTX 2.5 DEV merges bf16 / INT8 / fp8 - bring your own distilled LoRA

JoyAI-Echo's performance on the LTX-2.5 dev transformer, with nothing baked on top. These are for people who want to apply their own distilled LoRA at their own strength - or run the model slow, as a dev model.

Not few-step on their own. At 8 steps and cfg 1 these render soft and wrong. You either add a distilled LoRA (below) and run the normal distilled schedule, or run a dev recipe. If you just want to render, take the ready-made few-step files instead - Joy-LTX 2.5 on Civitai, or on Hugging Face GGUF / comfy-native. Those are exactly these dev merges with the official distilled LoRA baked in at 0.5.

What is in the merge

JoyAI-Echo (a fine-tune of LTX-2.3) is the better actor - lip-sync, expression, a voice that stays put. LTX-2.5's transformer is shape-identical to 2.3's, so Echo's delta against 2.3-dev transplants cleanly onto 2.5-dev: 070T30 = 0.7 x the video attention / feed-forward delta, 0.3 x on the modulation tables (the default dose); 100T50 = 1.0 x / 0.5 x (livelier, hotter). Nothing retrained, no distillation applied - the file is dev + delta, full stop.

How to use them

• Your own distilled LoRA (the normal way). Put a Load LoRA (model only) node between the model loader and the sampler with the official ltx-2.5-22b-distilled-lora-450-bf16.safetensors (Lightricks/LTX-2.5 on Hugging Face, loras/), then run the distilled 8-step schedule at cfg 1. 0.4-0.6 is where the picture keeps its natural grade - measured on the same seed, 0.4 / 0.5 / 0.6 are indistinguishable by eye and all cleaner than 1.0; 1.0 is fully distilled and hotter. 0.5 is what the ready-made files use.

• Any other distill / style LoRA the same way. That is the point of this page: the dial is yours.

• As a dev model. 25-30 steps, the LTXV scheduler, cfg about 3. Slower and softer, and the Echo delta reads as extra guidance at cfg > 1 - expect it hotter than stock dev; keep cfg modest.

LoRA on the int8 file works in stock ComfyUI - it is how Lightricks ship LTX-2.5 dev + distilled LoRA. The ComfyUI-JoyLTX25 canvases (Civitai: Joy-LTX 2.5 workflows) take these files too - drop the LoRA node after the loader.

Files (one version per dose)

• LTX25dev-echoVid-<dose>-DiT-comfy-int8.safetensors - 21.5 GB. Stock Load Diffusion Model; fastest on RTX 50, works on RTX 30/40 (ComfyUI 0.32+).

• LTX25dev-echoVid-<dose>-DiT-comfy-fp8.safetensors - 21.5 GB. comfy fp8_e4m3fn, scaled.

• LTX25dev-echoVid-<dose>-DiT-bf16.safetensors - 42.0 GB. The master, for merging or your own quantisation (if the upload here is refused for size it is on Hugging Face: joyai-echo-ltx25-echoVid-dev).

No GGUF on this page on purpose: a LoRA on a GGUF weight is dequantised per step and slow; quantise after you have baked your LoRA, or use the ready-made GGUF page.

You also need (Lightricks/LTX-2.5 on Hugging Face)

• ltx-2.5-video-vae-bf16.safetensors and ltx-2.5-audio-vae-bf16.safetensors → models/vae/

• text encoder gemma4-12b-with-proj-ltx-2.5-comfy-int8-convrot.safetensors → models/text_encoders/

• the distilled LoRA above → models/loras/; the x2 latent upscaler if you run two passes.

Why this exists

The first Joy-LTX 2.5 build applied Echo's delta to the distilled 2.5 transformer and came out over-cooked - +16% saturation, hard contrast, blown highlights on any bright scene. Diagnosing it showed the delta behaves like extra guidance on a few-step model, and that a dev-base merge with the distill LoRA held back to 0.5 reads clean. That number suits my eye; these files let you pick yours.

Credits

JoyAI-Echo by JD (jdopensource/JoyAI-Echo). LTX-2.5 and the distilled LoRA by Lightricks. Merge and files by me. Licensed under the LTX-2.x Community License (inherited from both parents). Questions: comment here or open a discussion on the Hugging Face page - I answer.