Master LongCat-Video-Avatar 1.5: 3 Pro ComfyUI Workflows |Whisper-Large + 8-Step DMD
Jun 3, 2026 #runninghub #LongCatVideoAvatar #LongCat1
Digital humans just reached a new level of stability! 🚀 In this video, we dive deep into LongCat-Video-Avatar 1.5, the latest update that significantly improves image quality, lip-sync accuracy, and temporal stability. Moving away from Wav2Vec2 to Whisper-Large v3, this model understands the nuances of speech rhythm across multiple languages (CN, EN, JP) better than ever. Plus, with DMD distillation, you can get high-quality results in just 8 steps! 🔥 What you'll learn in this tutorial: • 🧠 The 1.5 Evolution: Why this isn't just a minor patch—it's a complete overhaul of stability and clarity. • ⚙️ 3 Essential Workflows: Single: Perfect for short clips (under 10s). Extend: Manually chain segments for longer content. Auto-Extend: Automatically loops based on your audio duration. • 🛠️ The "4n+1" Rule: Avoid sampling errors by learning the specific frame math required for LongCat. • ✍️ Pro Prompting: Learn why audio isn't everything—how to use text to control body language and hand gestures. #runninghub #LongCatVideoAvatar #LongCat1.5 #DigitalHuman #LipSyncAI #ComfyUI #WhisperLargeV3 #WanVideo #AIWorkflow #AIVideo #TalkingHead #Kijai #DMDDistillation #AI数字人 #口型同步 #工作流分享 #人工智能 #数字人教程 ------------------------------------------------------------------------------------------------------------------------ resource model download： https://huggingface.co/Kijai/WanVideo... https://github.com/kijai/ComfyUI-WanV... online workflow 1：longcat avartart 1.5-single https://www.runninghub.ai/post/206146... 2：longcat avartart 1.5-extend https://www.runninghub.ai/post/206162... 3：longcat avartart 1.5-auto-extend https://www.runninghub.ai/post/206145... Runninghub Fan Benefit: Register to get 1,000 RHB https://www.runninghub.ai/?utm_source... workflow and prompt engineering download： https://github.com/amao2001/ganloss-l...
Transcript
0:00 · Let’s first look at a video This example is actually pretty interesting It shows a full-body view of a beautiful woman in traditional costume and the face does not take up many pixels in the frame The video resolution itself is not very high either
0:19 · only four hundred eighty by eight hundred thirty-two But if you look closely at the clarity of the face especially the lip movement the edges of the facial features and the overall stability of the character all of these are quite strong I think this is the most direct change in LongCat-Video-Avatar one point five
0:37 · It is not just simple lip sync or just making the mouth move The whole image becomes more stable cleaner and more like a digital human model that you can actually use for content creation Now let’s look at a few more examples from this model First it supports different languages
1:21 · Here, we have Chinese English, and Japanese You’ll notice that its adaptation to multiple languages is pretty good And it’s not only about normal speech When music and voice are mixed together it can still maintain relatively natural mouth shapes and expressions
1:38 · Of course that does not mean it has no problems We’ll talk about that later For example if the motion is not described clearly it can still look a little conservative So let me give you the conclusion first If you used the one point zero version before then version one point five is not just a small patch or minor update It has very obvious improvements in image quality lip sync, stability and usability for longer videos Officially they also mentioned that the audio encoder has been upgraded from Wav2Vec2 to Whisper-Large
2:10 · You can simply understand this as the model now having a more detailed understanding of pronunciation rhythm multilingual cadence and the mouth shapes corresponding to each phoneme At the same time it also uses DMD distillation to compress the generation process down to eight steps So while the quality improves the inference efficiency is also more suitable for real-world use So here comes the question: can we use this model inside ComfyUI?
2:37 · The answer is, of course, yes And today we won’t just look at the results We’ll directly break down three workflows for you The first one is called the single workflow It is suitable for short audio and one sampling pass can directly produce the result The second one is called extend It is suitable for long audio and you can manually extend the video segment by segment The third one I call auto extend It can automatically loop and extend according to the length of the audio
3:05 · After watching this video you’ll basically know how to use LongCat Avatar one point five inside ComfyUI and where things are most likely to go wrong Let’s start with the most basic part: installation and model preparation Here we are still using Kijai’s WanVideo extension If you have used the one point zero LongCat Avatar model before
3:25 · then this step should feel familiar What we really need to pay attention to here are the model files There are two very important files: one is the main model file and the second is an acceleration LoRA Both of these are actually new models The main audio model has also been switched We replace the traditional Wav2Vec2 with Whisper-Large v3 Right now Kijai has not officially released the one point five workflow yet so I will provide three versions: the single version the extend version and the auto extend version
4:19 · as we just mentioned Now let’s start with the most basic workflow What you’re seeing now is the single version which is the simple workflow Its basic structure can be divided into six parts The first part is model loading Here we load the BF16 main model for LongCat Avatar one point five
4:36 · At the same time we load an acceleration LoRA The LoRA weight is set to one point zero For the scheduler we use longcat_distill_euler Pay attention here: the step count is eight steps This is the eight-step distilled model emphasized by the official release The second part is reference image processing Here we scale the reference image to four hundred eighty by eight hundred thirty-two
5:00 · This resolution is not absolutely fixed You can adjust it based on the aspect ratio of your own image After the reference image is processed it goes into WanVideoEncode After this node it becomes the image embedding used during sampling You can understand it like this: the reference image is not simply pasted into the video It becomes a kind of condition that guides the model to maintain the character’s identity composition and overall visual style
5:39 · Then we use the Whisper model to encode the audio and it enters a very important node called LongCat Avatar Whisper Embeds This node is actually the most important step in the one point five workflow It converts the audio into audio embeddings that LongCat Avatar can understand
5:57 · In a way this works similarly to text embeddings The fourth part is the prompt In the single workflow I wrote a very simple prompt: “woman is talking happily ” The negative prompt is mostly regular quality control such as blur overexposure low quality, and so on Here I want to remind you that LongCat Avatar is still essentially driven by audio to generate video
6:20 · Its strongest ability is lip sync and digital human expression not complex action choreography So if you only write a very simple prompt like what I wrote here it will most likely give you a relatively stable and clean talking video
6:36 · But if you want obvious hand gestures leaning forward turning the head, smiling waving or even a little handheld camera movement then you need to clearly write those actions into the prompt Later we’ll do a dedicated comparison to emphasize how prompts should be written The fifth part is frame count processing There is a very important little trap here The frame count must satisfy the rule of being a multiple of four plus one
7:01 · For example, ninety-three one hundred forty-nine and one hundred seventy-three all work But if you randomly enter a frame count you may run into errors later during sampling So in this workflow I added a special handling step We first calculate how many frames are needed based on the audio duration and then use a formula to convert it into a multiple of four plus one This step looks small but it is very practical because the most annoying part of audio-driven video is that the audio length is not fixed
7:32 · You cannot manually calculate the frame count every time If the workflow can handle this automatically it saves a lot of trouble later The sixth part is sampling and output After sampling we decode through the VAE and generate a video Now let’s look at the result from the single workflow This Japanese example is about five point nine six seconds long
7:52 · with a total of one hundred forty-nine frames You can see that the lip sync expression and image stability are already very good So when your audio is relatively short for example within five seconds seven seconds, or ten seconds I recommend using the single version and sampling it directly It is the simplest option and is less likely to run into problems
8:14 · But in many cases when we make digital human videos we are not only making five or six seconds For example you may want to make a fifteen-second talking segment or even thirty seconds or one minute At that point one single sampling pass is not very realistic That is where the extend workflow comes in The idea is very simple We first generate the first video segment Then we take part of the ending of that first segment as the reference for the next sampling step
8:42 · By continuing segment by segment we can get a longer video In this workflow the biggest difference from the single version is that there are two very important sets of parameters: frames and overlap These also exist in the single workflow but they are not very important there You can basically set them however you want In this extend workflow I set frames to ninety-three and overlap to thirteen
9:06 · That means each sampling pass generates ninety-three frames but when two video segments are stitched together thirteen frames overlap So what is the purpose of overlap?
9:16 · It is not to make the video longer It is to create a smoother transition between two segments Imagine if the first video segment ends directly and the second video segment starts directly It is very easy to get an obvious jump in the middle The character’s face, hands hair or background may suddenly change So we keep a small overlapping region During the second sampling pass we use the last thirteen frames of the previous video as a reference
9:44 · allowing the model to continue generating from that state But there is a very important detail here During the final stitching for the overlap part we directly keep the latter part of the previous segment instead of keeping the beginning of the next segment Why?
9:58 · Because although the beginning of the next segment overlaps with the previous one it is still newly generated by the model Its quality is often not as stable as the previous segment So the safer approach is to keep the ending of the previous segment and then attach the next segment after the overlap In the workflow we use the ImageBatchExtendWithOverlap node to handle this stitching
10:20 · Its strategy is set to cut The advantage of this approach is that it is simple stable and less likely to go wrong Of course the problem with the extend version is also obvious: it requires manual extension You need to calculate and judge how much audio is left and you need to decide whether to run the next segment For testing, this is fine But if you want to batch process long videos it becomes a bit troublesome So we use the third version called auto extend The core of auto extend is basically adding a loop on top of the extend workflow
10:53 · Look here: it adds easy forLoopStart and easy forLoopEnd The workflow first calculates the total frame count based on the audio length Then based on the frames and overlap for each sampling pass it calculates how many loops are needed The formula used here is ceil(a / b), then minus one If the audio is very short and only needs one sampling pass then it does not need to enter the loop stage But if the audio is longer it will run several more rounds and attach the later parts So what scenario is auto extend best for?
11:25 · It is best when you have already tested all the parameters and made them stable and now you want to batch process audio that is ten-plus seconds or even dozens of seconds long Next let’s talk about a problem that is very easy to overlook: prompts When many people see an audio-driven digital human model they instinctively think the audio determines everything
11:47 · That statement is only half right Audio does determine lip sync speaking rhythm and part of the expression changes But the character’s main body movement the size of the motion whether the camera has a handheld feeling whether the character leans forward waves, or smiles all of these still need to be controlled through prompts Let’s look at two English examples In the first prompt the writing is relatively simple so the character basically just talks steadily The advantage is that it is clean and less likely to fail
12:18 · But the downside is also obvious: the overall movement is quite limited In the second prompt we describe the actions in much more detail For example the character starts with her hand behind her head then suddenly lowers her hand the corners of her mouth lift her body leans closer to the camera and she makes an open hand gesture The camera also simulates a slight handheld phone-camera shake
12:42 · When you look at the final result the motion becomes much richer This is an important usage trick for version one point five If you only want stable lip sync the prompt can be simple But if you want the video to be more expressive you need to write out the actions as much as possible and write them in chronological order Of course you must keep the bigger premise in mind: LongCat Avatar is still mainly a digital human model
13:07 · So I do not recommend asking the character to dance spin around or move quickly That is not what it is best at Alright, that’s all for today So what are you waiting for?
13:18 · Go try it yourself Follow me and become someone who truly understands AI