Request for official ComfyUI workflow and recommended parameters for JoyAI-Echo / LTX 2 & LTX 2.3

[Songssx]

Hi JoyAI-Echo team,

Thank you for open-sourcing this project. I noticed that the JoyAI-Echo model can technically be used inside ComfyUI now, because ComfyUI already supports LTX 2 and LTX 2.3 models.

I have tested it locally in ComfyUI. The workflow can run successfully and produce video results, but the final visual quality is currently not very good. I suspect this may be because the parameters used in ComfyUI are not fully aligned with the official recommended settings, and some details may need to be adjusted specifically for JoyAI-Echo.

So I would like to ask:

Could the official team provide a recommended ComfyUI workflow for JoyAI-Echo?
Could you share the recommended parameter settings for ComfyUI, such as resolution, frame count, sampler, steps, CFG, noise settings, conditioning settings, and other important parameters?
Are there any special settings or differences compared with the official inference pipeline?
Is it possible to provide an official long-video generation workflow for ComfyUI?

Since ComfyUI already supports LTX 2 / LTX 2.3, I believe an official JoyAI-Echo ComfyUI workflow would be very helpful for users who want to test and use the model more easily.

At the moment, my ComfyUI workflow can run, but the generated result is far from the quality shown in the official examples. It would be great if the team could provide an official reference workflow or parameter guide so that the community can reproduce better results.

Thanks again for the great work!

[Jahnsonblack]

Hi, thank you for your interest in JoyAI-Echo and for your feedback.

JoyAI-Echo currently requires the official inference script and configuration provided in this repository. Its generation quality depends on our DMD-distilled few-step inference settings and the paired audio-video memory mechanism used in the multi-shot long-video pipeline. Therefore, directly using a generic LTX 2 / LTX 2.3 ComfyUI workflow or default ComfyUI parameters may not reproduce the results shown in our official examples.

At the moment, we recommend using the official inference entrypoint:

python inference.py --config configs/inference.yaml

The key recommended settings are:

• Resolution: 1280 × 736
• Frames: 241
• FPS: 25
• dtype: bfloat16
• DMD denoising steps: [1000, 994, 988, 981, 975, 909, 725, 422, 0]
• DMD denoising sigmas: [1.0, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875, 0.0]
• Audio-video memory: enabled
• Memory max size: 7
• Number of fixed memory frames: 3
• v2a_grad_scale: 2.0

We understand that an official ComfyUI workflow would be useful for the community, and we will consider supporting and releasing a JoyAI-Echo ComfyUI workflow in the future.

Thanks again for your feedback and support!

[MeiYi-dev]

Hi, thank you for your interest in JoyAI-Echo and for your feedback.

JoyAI-Echo currently requires the official inference script and configuration provided in this repository. Its generation quality depends on our DMD-distilled few-step inference settings and the paired audio-video memory mechanism used in the multi-shot long-video pipeline. Therefore, directly using a generic LTX 2 / LTX 2.3 ComfyUI workflow or default ComfyUI parameters may not reproduce the results shown in our official examples.

At the moment, we recommend using the official inference entrypoint:

python inference.py --config configs/inference.yaml

The key recommended settings are:

• Resolution: 1280 × 736
• Frames: 241
• FPS: 25
• dtype: bfloat16
• DMD denoising steps: [1000, 994, 988, 981, 975, 909, 725, 422, 0]
• DMD denoising sigmas: [1.0, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875, 0.0]
• Audio-video memory: enabled
• Memory max size: 7
• Number of fixed memory frames: 3
• v2a_grad_scale: 2.0

We understand that an official ComfyUI workflow would be useful for the community, and we will consider supporting and releasing a JoyAI-Echo ComfyUI workflow in the future.

Thanks again for your feedback and support!

Can we get the base undistilled model as well please?

[Jahnsonblack]

Thanks for your interest!

The base undistilled model is planned to be released in the future. It is currently under internal company review, and we will update the repository once the release process is completed.

Thanks for your patience and support!

[merermero]

Proper Comfyui support with LTX2.3 is necessary. I can't test this.
I used my own comfyui ltx2.3 workflow this this model, it's not working well especially upscaler.
Thank you for your devotion!

[zhuang2002]

@Jahnsonblack @merermero @Songssx @MeiYi-dev
Hi all, ComfyUI support for JoyAI-Echo is now available.
You can use our custom node here: https://github.com/zhuang2002/ComfyUI_JoyAI_Echo

This implementation includes complete DMD denoising settings, audio-video memory mechanism and correct parameters, which can reproduce the standard video quality perfectly.

[zwukong]

we need cpu offload and gguf support. Native comfyui have both, hope someone can make it

[vskygk]

Can this run on a 32GB RTX 5090? If not, I guess I’ll have to wait for the FP8 or GGUF version.