Thanks Arkinson. I will have a look in the next few days and see if I can help.

@arkinson Wow. Thank you. This is awesome. I can't wait to try. I noticed previously that if my dialogue in vibevoice read like a song it had a really difficult time processing. Doing it this way might make it sound a lot better. Great stuff.!!

@arkinson I upgraded to .10 comfyui yesterday. It was a lot easier than I expected. I want to run Flux -2 Klein. It looks impressive.  

@arkinson I actually don't know. It took forever to load all the nodes from a workflow. I haven't checked it out this morning(for me). I need to test it and see. Development is so fast. I think Klein is the cut down Flux 2 model, which was too big for us to be really useable. I never got into Flux except for Inpainting and fixing hands. Flux excels at hand fixing.

@arkinson I am trying to run your new workflow but I am not sure where i should be getting. Do I get the ACE-Step file from here? https://huggingface.co/ACE-Step/ACE-Step-v1-3.5B I can't see the safetensors file.

@arkinson Nevermind. I think I found it. https://huggingface.co/Comfy-Org/ACE-Step_ComfyUI_repackaged/tree/main/all_in_one

@arkinson you are so quick!!! I am still generating on your workflow atm. I am interested in what Klein can do with consistent character, inpaints, face swaps etc. The fun never ends!!!!

@arkinson Interesting findings. Thanks for your run down. I have been playing with it. Using other peoples workflows. I used a really simple one and the results are very promising. I generated a simple gym scene and then placed the character from image 1 into the image 2 gym and the results were quite good without loras. It even kept body dimensions relatively the same. I tried another workflow to generate a multishot of 1 character, front, back, side and it had very mixed results. The Qwen results were much better. Possibly I didn't have enough steps. You are right, it is so quick. Great for just general images. It is still early days but definitely promising. I also have been trying to get to work a LTX2 workflows. Early days as well but looking a bit better. I need to play with it more. It also seems to be a Work in progress for the developer as well. https://civitai.com/models/2304098/ltx-2-19b-gguf-12gb-comfyui-workflows-5-total-t2vi2vv2via2vta2v

@arkinson Let me know how you go. I was trying to get the LTX2 Image with Audio to work (similar to your sound audio clip workflow) to work but I got a static image with the sound clip on top of it. The general I2V works ok, I need to keep testing. It looks very complicated and I think the developer is going to make some documentation hopefully.

@arkinson Thanks for checking it out Arkinson. I'm glad i'm not the only one getting static images, I tried changing a few settings and changing the prompts and it did seem to get a better response but I can't fathom exactly what it prefers. That is a super smooth continuation. I haven't tried this yet but is amazing what it can do.

@arkinson I think it is an audio issue. Basic I2V is fine for video., but sound was muted. Previously an I2V example gave me sound. It is way too advanced for me. I don't even understand how it all works. I think I will try some other peoples LTX2 Workflows to get a better understanding on how its put together.

@arkinson Someone helped me. Camera control Lora needs to be added. I tried this one and it worked like a charm. https://huggingface.co/Lightricks/LTX-2-19b-LoRA-Camera-Control-Static funnily enough its not on civitai but it does work. The other person used the dolly in. This helps a lot as we have a base line to work from.

@arkinson me again. Workflow author said this lora is good as well. https://huggingface.co/MachineDelusions/LTX-2_Image2Video_Adapter_LoRa/tree/main It also seems to do a good job, i dropped all strengths down to .5 that also included. ltx-2-19b-distilled-lora_resized_dynamic_fro09_avg_rank_175_fp8.safetensors as well as the ltx-2-19b-ic-lora-detailer.safetensors loras. I don't get the last image frame output though. Its a broken image icon. Do you get the last frame?

@arkinson i added a video helper suite select image node to get the last frame in the end. There are also new loras coming out for flux 2 Klein. I'm excited about this consistence lora https://civitai.com/models/1939453?modelVersionId=2634354 as even out of the box Flux 2 Klein is very accurate with body and face consistence somewhat.

@arkinson All models are the same I think the most important is to use the dev model gguf. ltx-2-19b-dev_Q4_K_M.gguf . I have 3 loras daisy changed off each other in the lora group at strength of 0.5. These loras in order. ltx-2-19b-distilled-lora_resized_dynamic_fro09_avg_rank_175_fp8.safetensors then ltx-2-19b-ic-lora-detailer.safetensors then LTX-2-Image2Vid-Adapter.safetensors all these loras are at strength 0.5 I did a 25 second clip it took 30 minutes. I am going to post it to the examples.

@arkinson  i dont know if you can download my workflow here. try it. https://limewire.com/d/tvAOj#SJWftpKge4

@arkinson I posted my example but there is always safety check delays. It will probably be there soon. 

@arkinson no worries. general i2v was no problem for me. It always seemed to work out of the box. Yes. taeltx_2 worked for me but it was being referenced in the easy comfyui folders not in my models folder. Even though i fixed the yaml it just wouldn't pick it up. I just copied the folder to the easy install folder and it worked. I couldn't be bothered in trying to make it perfectly referenced. I was surprised it went to 25 sec. I did it before i went to sleep and expect an OOM. It did it in 30 minutes. I will try longer and see how far it goes with either being too long or OOM.

@arkinson Excellent. I'm glad you got it to work. I can't believe how rapid the development is. Everyday there is some new model release, new workflow, new lora, new efficiency node. Damn!! its like Christmas never ends. I'm glad you are getting response with that audio package. There was a new Qwen TTS released a day or so ago. I couldn't load it, but people reckon the voice lacked emotion. Zimage base looks like it released as well. I 've only played with it a little.

I look forward to your version of the LTXv2 workflow. You seem to be able to get the maximum juice out of the model for us 12GB people.

@arkinson Your workflows are not rubbish now. They are still useful especially since Wan has a lot of NSFW loras behind it. It will take a while for LTX2 to catch up. Ltx 2 has got a lot of potential as long as its easy to make loras it will remain a strong contender in the future. But there are always new competition, I just wonder if it will ever slow down or it gets to the stage where we can literally dictate and control exactly what we want to see.

@arkinson I was trying to get frames to line up. there seems to be a delay in voice matching especially if you cut off mid sentence. Now that we can go to 20+ seconds input I don't think it will be a problem as I just need to cut at natural language pauses. I haven't test extensively lately, I am going on holiday soon and preparing. I wont be able to use Comfyui let alone a computer. So I will be gone for 2 weeks from Monday. I wonder what will happen in the AI field when I'm gone?

@arkinson Ha. Ha. not as bad as that but yeah its the things we do to keep the wife happy. I'm going to New Zealand to check out the north Island. Lots of driving, eating, sight seeing, very relaxed, just no computer . I looked up Goose Bay Labrador , wow that looks remote. I have never been to Canada. Thanks, I will talk soon.

@arkinson I just got back . We had a marvelous time at NZ. I think If I could, I'd swap living in Australia to living in New Zealand at times. But they are both similar countries. I can see you are doing great work again. I knew you would make an Arkinson special tweaked LTX2. I haven't tried it out yet but I will over the weekend.

@arkinson It was, thank you. One day you may find yourself travelling far. New Zealand and Australia are the perfect travel destinations. So much scenery, places, buildings, culture. I took lots of shots of the countryside on my potato phone camera. I am now in the process of using qwen image with those photos and having my usual fun. I am then going to put the images through your new LTX2 workflow. I will post results and give you any feedback that I might have.

@arkinson So how's life in your half of the equator? Are you working on any new AI projects?

@arkinson sorry to hear about that, I hope things get better. sounds interesting about your new endeavours. First frame last frame on LTX would be useful. I look forward to see what you produce. Over the last week developers have released a new Ace Step. for song generation. https://huggingface.co/ACE-Step/Ace-Step1.5 Apparently it is getting better over time and someone did a merge of the turbo model and base model. https://huggingface.co/Aryanne/acestep-v15-test-merges/blob/main/acestep_v1.5_merge_sft_turbo_ta_0.5.safetensors . I haven't tested yet but its on my to do list for AI. The big corps went for the jugular on music generation killing Udio and now Suno is also seeing big changes. So the only real solution is to generate locally. It can only get better.

@arkinson check out this hyperlink for some of my doctored travel pictures. https://civitai.com/posts/26581432 I am not sure if the link works for you as my pictures aren't coming up on the gallery. I used this qwen workflow https://civitai.com/models/2386462?modelVersionId=2683492 and the consistency lora. I honestly don't even need my Pony lora to generate simple posed images.

@arkinson I honestly havent tried T2V on Qwen as I am fixated atm on consistency. The tools have now gotten so good that you literally could take 1 photo and get a consistent SFW model pose from just that 1 photo. No LORA necessary. The workflow uses this massive AIO checkpoint that surprisingly works really well on 12gb VRAM just a bit slow to get an output. But since I am then using that image as a source image for LTXV2 video its worth it.

NSFW I think is more challenging as you won't get consistent nipples for examples when she disrobes. Hence the need for a chracter lora. Yes. She gets around. I just now need to find a distinct unique voice for her and then she is complete. I will create a back story for her, flesh out her personality and I will have her as a main character. As to the photos, they were real life photos without anyone in it. Its actually easier to slip a character into an empty shot than it is to prompt that qwen workflow to say remove the person in image 2 and place the person in image 1. That last photo where she is sitting at a table about to eat the Wendy's burger was actually me sitting at the table. I couldn't replace myself in the workflow, so I just said remove the person in the photo and I was gone. I then used that empty shot and prompted the person in image 1 is sitting behind the table in image 2. Viola. It worked. Although it reduced her boobs a bit. Anyway. It's all fun and games at the moment. I then had her speaking a few lines and eating the burger with your LTX workflow. It worked well for 2 tries and then it slowed down. So I have to work out what went wrong. The workflow is very solid. Yes a last frame image would be so useful.

@arkinson Ok. So I spent the whole day, enjoying the fruits of your workflow. Specifically I2V. a few things to keep in mind. If you prompt girl, you will get a high pitch almost kid like voice. see the example with my girl with the hat on pointing to the sign above her. She even acts a lot more kid like when you write girl. As soon as I refer to her as young woman, her gait changes, her demeanor changes and obviously the voice changes.

Ltx2V has problems with characters turning. Literally had to retry multiple times to get her to turn without body horror morphing. I feel that if you give the generation more time, it can turn her body a lot easier.

Natural speech has pauses between statements. So I just go .... between sentences to give it some speech pauses. The last generation was a doozy. LTXV2 does not naturally know what a penis is. I needed a LTXV2 Penis lora. Again I had massive problems with her turning around and then turning back to the camera. I think there is a LTXV2 prompt structure that I should be respecting. so That it is easier for the model to accurately reflect the users wishes. Anyway. It is a great workflow as always. Thank you for making it. I will play with the music version tomorrow. talk soon.

@arkinson Yes, I heard about the banning of certain images on Civitiai. Is it because there is no prompt ?

LTXV2 I think is still a bit early in LORA and workflow development. Wan has Wan animate that allows you to mimic the actions of a character and replace it with a character of your own choosing. Great for Tik Tok dance type stuff. I have looked but I cannot find this on LTXV2 yet.

It's great that people use your tools and create some funny stuff. Do you have a link to the reddit talking dog? I haven't tried V2V yet. Do you think it's possible to get an output of I2V from LTX2 and then feed that in as an input for a V2V? I wonder if you get some crazy town type stuff happening. I need to test more.

@arkinson Ohh I remember seeing this video when I was on holidays. It is very funny. I had no idea it was from your workflow. Good job. Crazy how good we can get on limited local resources. Your kiss video is great. I wonder how long we can keep generating from small snippets of video or if the video fidelity starts to break down over time.?

I haven't had any banning yet, I guess woman with big chest aren't considered children. If I did images of women with flat chest, perhaps the ban hammer might be enforced. I know they haven't got an AI bot for sound yet, as my swearing girl WAN video is still up.

@arkinson I have been playing around with LTXV2 more I can see what you mean about the distortion, especially with fast movement such as a dance. I generated a few I2V videos and prompted awkward dance and the last frame of the face is quite terrible. I will add it to the gallery for you to see. I tried to then clean it up by putting the last frame into a qwen workflow so that I could then use it again for another go but its not really a satisfactory output. I'm pretty sure its because its fast action.

Throw me a link to the lora manager. ? I might be using it already, I have a lora manager but I am not sure if I am getting full potential out of it.

I was wondering if I could show you a workflow and ask your opinion on how they got it to work? It is a Wan animate workflow that can go over a minute. It's good but I am not sure how they managed to get it to work for so long and the output is fair to average. If you have time to look. It does work on 12gb Vram .https://civitai.com/models/2018097/wan-animate-v2-unlimited-duration

@arkinson my videos wont go to the gallery. Must be because copyright songs. https://civitai.com/posts/26655378 is this the maximum output for movement we can expect on our 12gb vram cards? i am going to try some movement type loras and see if it improves. i think its because of the speed of movement of the character. especially hands. Its too fast.

@arkinson Yeah. further to my own understanding, LTXV2 still has limitations even if you had a large VRAM heavy card. This was a reddit post, I didn't read. Output is fair even though its from a RTX 6000 with 96Gb. https://www.reddit.com/r/StableDiffusion/comments/1q9cy02/ltx2_i2v_quality_is_much_better_at_higher/

I think I just have to be patient and not get too greedy and curb my over enthusiastic expectations.

@arkinson all good. I like looking at all the new stuff and the new models and workflows. I was trying to run the video output through an upscaler but at the moment its garbage in garbage out. I haven't seen you post at these hours of the day/night. Must be early for you.

@arkinson So, I put aside my quest for better I2V generations. I just installed this. https://civitai.com/models/2400306/ltx-2-easy-prompt-by-lora-daddy You might like it. basically it pimps your prompts, so if you are bad a prompt writing, it will breathe a bit of life into the prompt and supposedly give you a better output. It's definitely interesting and provides a lot of entertainment in what sort of generation you can get from the LTXV2 model. I stuck it into your workflow and works really well. I will post some examples in the gallery for reference. I just tested it more . It is giving me more variation in the output, mostly for the better.