Please be sure to update ComfyUI itself to the latest version.
v2.0 Prompt_Relay - Prompt-Relay has been implemented. Except for the positive prompt node, it is exactly the same as v2.0. If you want finer control over the prompts, this method is recommended, though it cannot be guaranteed to work perfectly.
v2.0 - Support for GGUF models has been added, and some 10Eros-Nodes have been implemented. Additionally, Distilled LoRA can now be disabled to allow the use of alternatives such as DMD LoRA. In the input area, it is now easy to configure settings such as the upscaled resolution, volume, number of steps for the Basic Scheduler, and Distilled LoRA weight.
v1.2 - Fixed connections related to the "LTXVCropGuides" node. Due to a lack of understanding of the "LTXVCropGuides" node, the connection method in v1.1 was incorrect. This error may have caused artifacts after upscaling under certain conditions, so this version has been updated, but it is not guaranteed to fix the problem.
Also, some configurations have been revised.
I noticed this issue when I tried to introduce a certain node into a new workflow, and severe artifacts occurred in my environment as well. Because the node connection method is different, it is not guaranteed that this workflow will be fixed, but I am releasing it for now.
v1.1 - Fixed an issue where the target FPS was being applied even when RIFE was not being used. Please enable only either "RIFE..." or "Don't Use..." on the FastGroupBypass node. Also, corrected an incorrect connection to the NAG node.
v1.0 - I've released the workflow I've been using for generation. However, please don't expect too much support from me.
●The safetensors model required for this workflow:
【Text Encoder】
coomfyUI/models/text_encoders
gemma_3_12B_it_fp8_scaled.safetensors （Some people also use fp4.）
https://huggingface.co/Comfy-Org/ltx-2/tree/main/split_files/text_encoders
or
Gemma_3_12B_it_ablit-norms-biproj-fp8mixed.safetensors
https://huggingface.co/TenStrip/LTX2.3-10Eros/tree/main/text_encoders
or
gemma-3-12b-it-heretic-v2_fp8_e4m3fn.safetensors
ltx_text_projection_bf16.safetensors
https://huggingface.co/Kijai/LTX2.3_comfy/tree/main/text_encoders
(If using checkpoints that include CLIP, this is unnecessary.)
【VAE】
coomfyUI/models/vae
LTX23_audio_vae_bf16.safetensors
LTX23_video_vae_bf16.safetensors
(If using checkpoints that include VAE, this is unnecessary.)
taeltx2_3.safetensors
https://huggingface.co/Kijai/LTX2.3_comfy/tree/main/vae
【Upscaler】(Be sure to select "1.1". "1.0" has a bug.)
ComfyUI/models/latent_upscale_models
Ltx-2.3-spatial-upscaler-x2-1.1.safetensors
https://huggingface.co/Lightricks/LTX-2.3/tree/main
●Recommended safetensors models for this workflow:
【Recommended Checkpoint】
(include CLIP & VAE version) comfyUI/models/checkpoint
ltx2310eros_v12.safetensors
ltx2310eros_v14.safetensors
https://civarchive.com/models/2447875/ltx23-10eros?modelVersionId=3109610
or
comfyUI/models/diffusion_models (GGUF … comfyUI/models/unet)
DasiwaLTX23_DragonleapV4.safetensors
https://civarchive.com/models/2543443/dasiwa-ltx-23?modelVersionId=3092188
【Distilled Lora】
ComfyUI/models/loras
Ltx-2.3-22b-distilled-lora-1.1_fro90_ceil72_condsafe.safetensors
https://huggingface.co/TenStrip/LTX2.3_Distilled_Lora_1.1_Experiments/tree/main
※ Recommended weights for Distilled LoRA
・ 10Eros v1.2 ... First Pass: 0.8–0.9, Upscale Pass: 0.5–0.6
・ Dasiwa Ltx2.3 v4 ... First Pass: 0.5–1.0, Upscale Pass: 0–0.5
Distilled LoRA and DMD LoRA serve similar functions, so do not enable them at the same time.
If the base model recommends DMD LoRA (such as 10Eros v1.4), enable only DMD LoRA and disable Distilled LoRA. For other base models, enable only Distilled LoRA and disable DMD LoRA.
【DMD Lora】
ComfyUI/models/loras
LTX2.3_DMD_reshaped_r256.safetensors
https://huggingface.co/TenStrip/LTX2.3_DMD_Lora/tree/main
※ Recommended weights for Distilled LoRA
・ 10Eros v1.4 ... 1.0
【OmniNFT Lora】
LTX-2.3-OmniNFT-RL-Lora_bf16.safetensors
https://huggingface.co/Kijai/LTX2.3_comfy/tree/main/loras
【comfy-nodes】
https://github.com/TenStrip/10S-Comfy-nodes
【Prompt-Relay】
https://github.com/kijai/ComfyUI-PromptRelay
Even when using the fp8 model, downloading all the necessary files will exceed 40GB. Because of this, you will likely encounter an Out of Memory error unless you have at least 64GB of system RAM. In fact, I hit OOM errors myself depending on the resolution and video length I tried.
Although using Prompt-Relay makes the workflow somewhat more complex, it offers significant benefits by enabling time-based prompting. It is a shame that it doesn't allow for dynamic adjustment of LoRA application—unlike the SVI feature in Wan2.2—but I consider it an essential node if you want to create videos that more faithfully match your vision.
Description
v1.2 - Fixed connections related to the "LTXVCropGuides" node. Due to a lack of understanding of the "LTXVCropGuides" node, the connection method in v1.1 was incorrect. This error may have caused artifacts after upscaling under certain conditions, so this version has been updated, but it is not guaranteed to fix the problem.
Also, some configurations have been revised.
FAQ
Comments (28)
Text to video option? :)
I have never tried text-to-video before, so please look elsewhere.
@kenpechi So far it's great workflow. i look forwaards for updated version, i add Olama.. and got great result!
thank you, love the workflow, any way to improve facial adherence to original image?
From what I've tried, this is a fundamental issue with Ltx2.3. I try to eliminate LORA as much as possible, but sometimes faces still look like different people. The node developed by the creator of 10Eros might mitigate it somewhat, but I haven't released it yet because I haven't seen a clear benefit.
@kenpechi thank you for your response, regardless your workflow has been the best one I've used for i2v, hope to see more from you!
dev-transformer-fp8+distilled lora 0.42+multi-concept 0.3+nsfw_furry 0.3+better motion 0.4
Replacing the image with a 5-second HD video clip inserted into Wan2.2 will improve the effect; I use this method for my LTX2.3 videos.
All of your work is amazing, my friend. Thank you
Been testing out your workflow all morning and its been fantastic! Good Job. Would love if you could incorporate voice clone into it.
Just wanted to say thank you for sharing.
Keep up the good work 👍
The best workflow I've seen on LTX. True, version 1.1 had a lot of artifacts. Version 1.2 generates perfectly, without artifacts, but I encountered a problem where the RAM from the previous generation isn't cleared during a new generation, causing an insufficient RAM error. And until I restart Comfyui, the errors keep popping up. Is there a way to fix this?
Replacing the two VAE Decode nodes in the First Pass and Upscale Pass subgraphs with "VAE Decode(Tiled)" should significantly reduce the risk of OutOfMemory (OOM).
There’s been a good opportunity; it would be nice to screw that thing in there. LTX Director
The Ltx Director workflow is already complete and will be released soon. However, it basically follows the creator's workflow, so it won't be anything particularly groundbreaking.
@kenpechi will you be using ltx director 2.0?
@rogerea421 Yes. However, since we've added more features, we'll organize them before releasing it.
When I ran your workflow earlier, I encountered this error: LTXVImgToVideoInplaceKJ AttributeError: 'NoneType' object has no attribute 'downscale_index_formula. The left input of the Patch Sage Attention KJ node is completely unconnected. I didn’t edit anything in the workflow; it was already disconnected as soon as I downloaded it. I have loaded all the required models correctly, but I still can’t identify the root cause of the problem. If possible, could you upload this workflow to Running Hub? I’d like to review it and figure out which connections need fixing. Thank you very much for your help!
I’ve just downloaded and checked it again, and there are no issues in my setup. Someone else mentioned the exact same thing, but if the connection on the left side of the "Patch Sage Attention" node is disconnected, please connect it yourself. Connect it to the purple "model" input on either the "Load Diffusion model" or "Load Checkpoint" node (located at the top left). My setup is configured to work with either model, but the bottom line is that it just needs to be connected to one of them. I have no idea why it’s disconnected, but I’ve verified that the data itself is fine.
There is no need for me to go through the hassle of uploading it to the Hub or anything like that; please resolve the issue yourself.
@kenpechi I finally managed to solve the issue. I'm not sure why, but several nodes were disconnected after I downloaded the workflow. Maybe there's some difference between the ComfyUI package I'm using and yours.
In the end, I found that the left input of the Patch Sage Attention KJ node wasn't connected to the model output of the checkpoint node. Also, the vae output of the checkpoint node wasn't connected to the Set Vid_VAE node.
I'm a beginner, so I really appreciate your patience and your replies. The workflow is great—thank you!
Instead of opening the workflow from the ComfyUI workflows folder, try clicking the logo in the top-left corner, selecting File > Open, and choosing the workflow from your own files. That works for me, because when I opened them directly from the workflows folder, the nodes I didn't have installed wouldn't show up.
10Eros V1.2 and 10S-Comfy-nodes V1.73 have been updated. When will the new workflow be released?
I’m currently focusing on ltx Director, so there likely won’t be any new releases for a while. While I still need to measure the effectiveness regarding the nodes, the 10Eros v1.2 model works without any issues.
I haven't found a better workflow yet. When I listened closely and heard it actually talking, I nearly fell off my chair.
Been testing your workflow for a few days now and it's probably the best I've used. However, I'm getting some issues with ghosting, specifically the eyes. Is there a setting or fix to avoid that?
The issue you pointed out is specific to Ltx2.3, and it appears that a fundamental solution is impossible. Various workarounds have been suggested—for instance, in discussions regarding 10Eros nodes—but none have yielded a clear difference.
Excellent workflow, thanks! The only problem i have is that i would like to set up the resolution automatically but i have to do it manually for each picture and when i try anything above 720p, the result is never good, looks stretched or some other problem usually. i tried to add it to the workflow it just messed it up so i just kept using it manually. To be fair i also had another workflow and even in that one my usual method doesn't really seem to work how i want it to work. I think the only one i have been able to make it work has been with the official LTX workflow.
This is a super cool WF, thanks for this. If I had any suggestion it would be to implement an optional Last Frame guidance. I was comparing the LTX2.3-FFLF WF from Comfy templates to yours and I can see that the extra features from KJ nodes adds quite a bit of complication. I think it would be worth the effort, if you ever had the ambition. I've been able to work complicated things into some other WF's but this seems over my head.