Please be sure to update ComfyUI itself to the latest version.
FMLF v2.0 – This version isolates the "First, Middle, and Last Frame" functionality from Director v2.0 and separates it into three distinct nodes. Since a number of users find the FMLF version easier to use, I have updated it alongside the main version.
Director v2.0 - Distilled LoRA can now be disabled to allow the use of alternatives such as DMD LoRA. I have made it possible to switch between the manual sigma value and the "Basic Scheduler."
※ Please note that the audio volume inevitably drops after upscaling. I am not sure if this is due to a mistake in my settings. For now, I have set the volume to "18," but if the volume is fine in your setup, please adjust it to "0" or another appropriate level.
Director v1.0 - I have introduced the "Ltx Director v2.0" node and removed nodes with redundant functionality. I also updated the input area to allow adjustment of the Basic Scheduler step count and the target resolution following the final upscaling. Additionally, I added nodes for configuring Ic LoRA settings.
FMLF v1.1 - Moved the settings for frame rate and length (seconds) loser to the Prompt-Relay node to improve usability.
Made it easy to apply frame rate changes to the Ltx Sequencer.
Consolidated the Power LoRA Loader nodes into a single node, as the downsides outweighed the benefits.
Made it easier to adjust the volume of the generated audio.
Switched the VAE Decode in the upscaling path to the Tiled version.
Expanded the list of recommended checkpoints after testing with models other than 10Eros.
FMLF V1.0 - This workflow incorporates two nodes by WhatDreamCost, "LTX Sequencer" and "Multi Image Loader," as well as Kijai's "Prompt-Relay" node.
Please refer to WhatDreamsCost's GitHub for information on how to use the "Ltx Director v2.0" node.
【WhatDreamsCost-ComfyUI】
https://github.com/WhatDreamsCost/WhatDreamsCost-ComfyUI
【Prompt-Relay】
https://github.com/kijai/ComfyUI-PromptRelay
This workflow is not suitable for beginners.
"Ltx Director," provided by WhatDreamCost, allows you to place multiple "anchor images" at specific points along the video timeline and assign corresponding prompts to each. This capability makes it easier to control the video's content by applying the necessary images and prompts exactly when desired. Furthermore, when featuring the same character, using their image as an anchor helps maintain facial consistency.
However, the key difference from standard workflows is the requirement to prepare multiple images in advance; consequently, this workflow may not be suitable for beginners. Personally, I primarily use "Qwen-Image-Edit 2511" to edit image content and orientation, while using image processing software to handle cropping and the preparation of multiple images. Although it requires some effort, once the necessary images are ready, you can create videos that faithfully reflect your creative vision.
Additionally, the "FMLF" version is a workflow that isolates the "First," "Middle," and "Last Frame" functions of Ltx Director into three separate nodes. If you prefer a more streamlined feature set to achieve greater flexibility in video control, you might find this version easier to work with.
I only recently began using this workflow in earnest, but I decided to share it now in the hope that it contributes to the further expansion of the local generation environment for Ltx2.3.
●The safetensors model required for this workflow:
【Text Encoder】
ComfyUI/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】
ComfyUI/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.)
Ltx-2.3-spatial-upscaler-x2-1.1.safetensors
https://huggingface.co/Lightricks/LTX-2.3/tree/main
●Recommended safetensors models for this workflow:
【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
【OmniNFT Lora】
LTX-2.3-OmniNFT-RL-Lora_bf16.safetensors
https://huggingface.co/Kijai/LTX2.3_comfy/tree/main/loras
●As we have received many inquiries, I am sharing the models and workflows used in Qwen Image Edit 2511 for your reference.
【Recommended Checkpoint】
Qwen-Rapid-AIO-NSFW-v19.safetensors
https://huggingface.co/Phr00t/Qwen-Image-Edit-Rapid-AIO/tree/main/v19
【Recommended Workflow】
https://civarchive.com/models/2167203/nsfw-qwen-image-edit-rapid-aio?modelVersionId=2440501
【Multi-angle LoRA model】
（It is already included in 2511, but using LoRA is expected to further enhance the effect.）
qwen-image-edit-2511-multiple-angles-lora.safetensors
https://huggingface.co/fal/Qwen-Image-Edit-2511-Multiple-Angles-LoRA/tree/main
【Multi-angle Workflow】
image_qwen_image_edit_2511_multiangle_camera.json
https://github.com/jtydhr88/ComfyUI-qwenmultiangle/tree/master/workflow
● Points to Note:
・ Even when using fp8 models, the total size of the required files exceeds 40GB. Consequently, unless your system has at least 64GB of RAM, you are highly likely to encounter Out-of-Memory (OOM) errors. Since this workflow supports GGUF models, you might want to consider using them instead.
・ When inputting multiple images, you need to adjust the "Guide Strength" value at the bottom right of the node for each segment. Although this setting is small and easily overlooked, it is crucial; please take the time to specify it for every segment. Note that a lower Strength value for an applied image allows for more movement during the video, though this comes at the cost of consistency regarding the anchor image. Therefore, it is best to fine-tune this setting based on the specific video you wish to create.
・ Ltx2.3 Distilled LoRAs are notoriously difficult to work with—even more so than those for Wan2.2. While many models have been released, it is hard to determine which one is best. You will need to experiment with various combinations to find what works with your chosen base model; the same applies to weight settings.
・ Although IC-LoRA functionality is included, I have not used it extensively. I did run generations using the models "ltx23_edit_anything_global_rank128_v1_9000steps_adamw.safetensors" and "ltx-2.3-22b-ic-lora-union-control-ref0.5.safetensors" to verify operation; while I cannot speak to the output quality, I confirmed that the IC-LoRA function was indeed applied. Please understand that I cannot answer questions unrelated to IC-LoRA workflows. The tool is highly feature-rich, so please refer to the creator's GitHub page for full details.
・ Just to clarify, I am not an expert in AI generation. Everything I’ve shared here is the result of trial and error, based on actual testing and research into the vast amount of information available in the community. Consequently, I cannot answer highly technical questions; my knowledge is strictly limited to my own personal experience. If any issues arise, I would greatly appreciate it if you could handle the troubleshooting yourself.
Description
Director v1.0 - I have introduced the "Ltx Director v2.0" node and removed nodes with redundant functionality. I also updated the input area to allow adjustment of the Basic Scheduler step count and the target resolution following the final upscaling, and Distilled LoRA weights. Additionally, I added nodes for configuring Ic LoRA settings.
FAQ
Comments (28)
Looks very promising, I'll try it out soon. I need a bit of help on how to create more 'similar' images based on 1 base-image though. Do you have any tips / recommendations for that?
What I’m doing is simply using the straightforward "Qwen image edit 2511" workflow to adjust angles and make partial corrections—such as fixing a skirt that has flipped up. I’m using a workflow found on Civitai. While it doesn't allow for major edits, it’s enough to add a nice touch to the video.
Those who know how to use face LoRAs could likely get even better results with this workflow.
For image cropping, I use GIMP; it’s free and highly versatile.
@kenpechi 可以告诉我你使用的是什么工作流吗？我找了几个Qwen2511的工作流并不是很理想。还有我想请问下你写提示词的思路是什么吗？
I have added the following text to each of the main bodies.
●As we have received many inquiries, I am sharing the models and workflows used in Qwen Image Edit 2511 for your reference.
【Recommended Checkpoint】
Qwen-Rapid-AIO-NSFW-v19.safetensors
https://huggingface.co/Phr00t/Qwen-Image-Edit-Rapid-AIO/tree/main/v19
【Recommended Workflow】
https://civitai.red/models/2167203/nsfw-qwen-image-edit-rapid-aio?modelVersionId=2440501
【Multi-angle LoRA model】
（It is already included in 2511, but using LoRA is expected to further enhance the effect.）
qwen-image-edit-2511-multiple-angles-lora.safetensors
https://huggingface.co/fal/Qwen-Image-Edit-2511-Multiple-Angles-LoRA/tree/main
【Multi-angle Workflow】
image_qwen_image_edit_2511_multiangle_camera.json
https://github.com/jtydhr88/ComfyUI-qwenmultiangle/tree/master/workflow
@kenpechi 谢谢
I might be doing something wrong; I add one image and basically instantly have a different face from the one on the image.
I encounter the same problem sometimes, dont know how to do to keep the exact face as input image, it always change a little and dont look like exactly as the original
What lora do you use? I found out that dreamlay for me completly changes the face
As you know, facial inconsistency is a typical weakness of Ltx2.3. Consistency varies depending on factors such as the desired action, the type of checkpoints used, and the resolution.
Naturally, the 10eros model offers the highest consistency, and a resolution of at least 720p is required.
If the face still becomes distorted, try inserting an anchor image just before the distortion occurs or changing the specified action.
This workflow is Amazing, great job. I just have a little default, input image have bad fidelity with face and change a lot after few seconds, do u have some recommandations ? Higher steps ? Better prompt ? Thanks !
As you know, facial inconsistency is a typical weakness of Ltx2.3. Consistency varies depending on factors such as the desired action, the type of checkpoints used, and the resolution.
Naturally, the 10eros model offers the highest consistency, and a resolution of at least 720p is required.
If the face still becomes distorted, try inserting an anchor image just before the distortion occurs or changing the specified action.
Very good workflow, works well! One thing, though: In the "Don't use RIFE"-Subgraph, the video combine node says "loop count 19" by default instead of 0. Took me a while to figure out while my video kept repeating for a few frames. Yet very simple solution.
Thank you for pointing that out!
I will swap the data to avoid any confusion.
i love it,but I wanna know use director2.0 to make a video will consume more time than the simple I2V ?5090 15min to make a video,Is this normal?
Using Director doesn't result in longer generation times. While the initial load takes a while, in my experience, it’s no different from I2V.
@kenpechi Thank you for your reply. With simple I2V, I can generate a 20-second video in just 5 minutes, it takes 20 minutes for the director to produce the same video
@woaoaj Generation times vary depending on various factors within the workflow—such as the Upscale sigma values or the use of Tiled VAE decoding—so it is difficult to give a definitive answer. If you are able to modify the workflow yourself, please change the sampler sigma values for both the First Pass and the Upscale Pass to match those used in your I2V workflow. I have left the nodes bypassed so that the sigma values can be set manually. Additionally, please switch the Upscale Pass VAE Decode to the standard version. Doing so should likely reduce the generation time. Please understand that the workflow was designed this way to ensure broad compatibility across a wide range of user setups.
I’m using your fmlf1.2，it works perfect and ！takes 5min to generate a 20-second video，thanks for your work！
wheres the download link for FMLF v1.2 ?
Sorry about that—I was swapping out the data to fix a minor error. Or perhaps it's invisible right now because the only video I've posted so far is undergoing moderator review? In any case, I'm uploading a new video, so things should go back to normal before long.
Kenpechi.......is King of Erotic. Wonderfull Work
Should allow_compile be enabled or kept disabled inside the Patch Sage Attention KJ Node? And should sage attention be on auto or can i select fp16 cuda? Im using LTX Director v1.0
I’ve never enabled "allow compile," and I don’t really have much knowledge about SageAttention to begin with, so I’m not sure.
What about try to use 10S Nodes for most final video quality boost?
Did you confirm that using it improves quality?
@kenpechi I test it now.
Well they don't boost quality, only thing worth adding is "reference memory" node and reference conditioning. Look at my DMD workflow to see how it goes in but I can explain in DMs if you want.
@tenstrip https://civitai.red/models/2731045/compact-ltx-23-eros-workflow-i2v-10s-nodes-stg-guider-rtx-upscale-rife-original-by-kenpechi - there is modified @kenpechi 's workflow, looks like it gives a boost quality picture and animation. Perhaps the issue is not only with these nodes but also with some other settings.