This is an experimental alternate distillation LoRA based on JoyEcho DMD for LTX 2.3.
Tested settings:
Strength
0.4-1.0Works with:
1st passand2nd passand3rd pass8-12steps
Limited testing was done. Good luck! ✌️
The Goal:
This LoRA recreates the distilled JoyEcho LTX 2.3 model changes on top of the original LTX 2.3 Dev checkpoint, so it keeps JoyEcho’s enhanced distilled behavior.
All standard model routes are included. Unlike an I2V CondSafe LoRA, nothing is intentionally removed.
V2
I raised the overall rank and did a selective SVD to get more quality.
Sound and artifacts are noticeable better on this one, from my testing.
The final audio-output route is blended: 20% of the JoyEcho change and 80% of the official LTX distilled LoRA change.
The LoRA is compressed dynamically to a maximum rank of 86 per layer while targeting 99% retained weight energy. And target 160 rank for Audio.
V1
The final audio-output route is blended: 25% of the JoyEcho change and 75% of the official LTX distilled LoRA change. This keeps the better distilled audio behavior from the official r384 while reducing the overly loud sound produced by the full JoyEcho output route.
The LoRA is compressed dynamically to a maximum rank of 72 per layer while targeting 96% retained weight energy. This makes it substantially smaller and more practical than a full checkpoint delta, while preserving the most important learned changes.
Description
V2
Audio Rank ~160
Global Rank ~86
Audio out path LTX2.3 80% / 20% DMD
FAQ
Comments (7)
is this needed for your latest LTX 2.3 model ?
No, not needed for DragonLeap v4
@Darksidewalker I was wondering whats this needed for ? Your older models or ...?
@rawrrawrrawr3x379 Older models and non-distilled models.
You say, 'This LoRA recreates the distilled JoyEcho LTX 2.3 model changes on top of the original LTX 2.3 Dev checkpoint, so it keeps JoyEcho’s enhanced distilled behavior.' This is completely incomprehensible and I have been working with Comfyui for 2 years, so imagine newbies. Keep your technical details for the specialists. What is the PURPOSE of this Lora ? Why should I install it compared to regular LTX workflows? Can you provide a workflow that uses it and give a comparison without ? Is this Lora work with I2V workflows ? Thank you.
I've been working with Comfy for 2 years and it makes perfect sense, have you tried other DMD LoRAs for other checkpoints? do you know what distillation means? look them up and you'll understand why that statement is quite informative, technical details matter. An extra “makes generation faster with better quality” would be helpful though.
Works well with Eros1.4 and 1.5 (1.5 on hugging face only currently). It does help to add movement in i2v. Videos generally look a lot less tiff than the default DMD lora. Mainly used 1.00 weights but I'll try 0.6-0.9 and see if I can squeeze more life out of it.
Tip: the 10s Node FaceID workflow helps a lot with faces too (even better after bypassing upscale pass).