same vision encoder as wan2.2, same text encoder as Qwen-Image + byt5_small_glyphxl, new vae

@sweetmax797 Thank you!

@sweetmax797 Sage attention question.
So im up and running, figuring out new quirks, You mentioned sage attention issue.
I run sage attention, so curious about that.
I downloaded the 480 T2V and it seems fine on sage, but there is 720p not distilled t2v listed on huggingface, is that the one with sage issue?
I had issues on original hyv with sage, but it was only fp16 cuda (black output). fp16 triton and and fp8 cuda were always fine

@mrreclusive3545 No, the Lightx2v 4-steps models. I tested all of the i2v and t2v models and encountered a black screen, even without Sage attention. However, with the distilled models, my setup had no issues, both with and without Sage. I guess it has to do with the version of PyTorch(2.9). you hadn't any issues with the Lightx2v models or jsut regular fp16?

@sweetmax797 Ok, so yes, I am seeing same thing, every lightx2v i download gives me black screen, with or without sage, (my pytorch is 2.7)
Are you going to post a 720 distilled for t2v, I only see 720 distilled for i2v right now.

@mrreclusive3545 There isn't any distilled FP8 for 720p (at least for now). As soon as it is out, I'll try to upload that too

@sweetmax797 Thanks for the info, and just an fyi, I am setting up a second comfy to see if I can get the 4 step to work, every step of the way I get weird dependency issues.
if I find the answere, I'll let you know.

@mrreclusive3545 Sorry for the late reply. It won't work! those 4-steps models currently only work with Lightx2v inference (https://github.com/ModelTC/LightX2V). As of now, there isn't any Lightning 4 or 8 steps LoRA or merge for ComfyUI.

@sweetmax797 no worries, and there is lightx stuff for comfy, but it all seems broken atm, wont finishing building the wheel.

in which workflow? it's private node I use, must be there by mistake, just use the core "Load Diffusion Model" node.

@sweetmax797 Thanks, I had same issue, i did switch to normal loader, right now waiting for vae/clips to download

i just talked to him about this, yes there is a 4step, but its not working for alot of us, so the standard distilled, which does work, "needs" 50 steps, some of my early testing though, with sage, and no cache, i get decent results at 12 steps with standard distilled

its in the video's, just drag the video into comfy.

workflows are in the zip files (each Training Data files contains a workflow )

yeah, its only been out for a day now so, things take time. ;)
But for i2v/t2v, isn't it just based on the model you train against? in a lot of cases the t2v and i2v lora's seem to be interchangable, not always though.
also through my own experience in 1.5, ive done 720p t2v by running the 480t2v model for a few steps, then letting the i2v model complete it (like running wan2.2 moe) and it does fully stabilize the latent.

@mrreclusive3545 i am curious if old hunyuan loras will work with it

@snap2887 nope, the model seems built completely different, none of the blocks line up. kind of like how wan lora's don't work between versions, or even variations like the single model and double model wan lora's won't mix. same thing there, internal structure is to different.

theoretically the block's could be "remapped" to the correct space, but even then, this new model may run so different that won't work.

@mrreclusive3545 great, its better to know there is no hope for my old Loras. glad that new ones are on the way.

@snap2887 I know, have so many lora's trained on hyv1, guess im going to delete them all.
things are moving fast though, we just got the lightx lora last night, so should have lora training soon.

@snap2887 im looking everyday for anything for lora training, even just talk of anyone working on one, nothing yet.
i did try to block reassign one of my lora's to see if i could get anything from hyv1 to work in 1.5, no luck.

My 2 cents: this is just 1 model with fewer parameters, so it should be much easier to train, distribute and use. Now that we have Lightx2v acceleration, Hunyuan 1.5 is faster and results in higher frame rate outputs. I'm looking forward to LORAs for this.

@Phr00t_ hey man now trying your all in one baked model with the acc... can you point me out how to learn to train a lora for this hunyuan 1.5 model? i just keep waiting for now for diffusion pipe\one trainer support, but got no clue how to do it :D

Everyone keep in mind this is about these release 1.5 models and i2v only. The I2V is simply terrible compared to Wan. LightX fixes the outputs alot but has a ton of new problems.

It's heavily censored compared to v1. That's for sure. I agree that v1 was better. This seems like a downgrade. For the 95% of the community that only cares about the naughty anyway 😅. We'll see what the community can do but it's a shame Hunyuan decided to censor their model. I had high hopes and now they're crushed.

@BigSad11 The Kadinsky5 model that will be around soon is gonna probably be a better alternate model to Wan. Out of the box it seems to have been somewhat tuned on porn it can render genitals and penetration T2V.

@tenstrip Thanks! First time hearing about this model. I'll have to check it out. Cheers.

@BigSad11 These are some base model Kadinsky5 Text to Video outputs I've seen shared https://postimg.cc/gallery/mhc4dTY

@tenstrip I hate to take up any more of your time but I was having issues downloading it. Something with the requirements. The vae and text encoder nodes didn't show up after installing. I've given up for now but will try again later. Did you experience anything like that?

@BigSad11 All I've seen is that you have to use this comfy fork because it's not supported officially yet https://github.com/kijai/ComfyUI/tree/kandinsky5

@BigSad11 i have tried the i2v 5lite but got bad results. when you manually downloading the custom nodes it you need to run "download_models.py" in custom nodes(to downloads the vae\text econders), ask claude\chatgpt to remove the base models(it is downloading them all) so it only download text encoder and vae, then you mv it (the whole folders) to Comfyui/models /vae or textencoder

i didn't tried t2v.. but this results is not bad

@BigSad11 i used teh wan2gp workflow with this model, and it had NO issues in continuing a 5 second porn clip in precisely the way i described, with full nudity and penetration. You just aren't using it right

@BigSad11 i did look into the development with this model, they state its not censored, just not trained on explicit content, so if/when we get lora/training, we should see some good nsfw content. I have the exact opposite issues with the wan2.2 models here on civit, getting it to NOT make nsfw content. lol

@flacid you're using wanGP with it? I can't even get a basic working comfyui workflow going with sft-pro-5s i2v model. Every output has a lot of noise in the first frames and is ruined and I can't find any workflows for it.

@tenstrip Also curious about this myself.

Distilled models are here due to the lack of Lightning LoRAs. They require less CFG and are usually quicker. As soon as Lightning LoRAs are released, I will upload them along with the Lightning LoRAs

@sweetmax797 im running the fp16 720 right now along side the fp8 480 distilled (running direct comparisons) speed is the same, cfg 1 on both.
the fp8 is giving "cleaner" output, but, is ignoring 90% of the prompt, the fp16 follows prompt a lot better.
but.
I am aware, I have a unique setup, most people can't load a 16gb model without blockswap, so yes, in that case, the the smaller model would be faster.

@mrreclusive3545 Of course, FP16 is better, there are no debates about it! The majority of people here have low VRAM, from 6 to 15 GB. FP8 is popular because you can run it and still do something else on the PC. I use FP16 or BF16 for most tasks, only testing FP8 occasionally. However, you could upload all FP16 models and I'll add your repo as "suggested Resources"

@sweetmax797 ill probably do that, just uploaded a bunch of comparisons too, im also trying to figure out how to fully utilize the 1080 SR model, I know its a standalone upscaler model, just not sure the best workflow yet.

@mrreclusive3545 Great! the workflow is in zip files, Dataset Files contains workflows,

I'm so sorry, I have to remove the Lightning LoRA. It seems the website now allows creating new repositories for models. I see you shared some videos, they will be removed too, please take it kindly.

@sweetmax797 no worries, reassigned them to the new lora ;)

@mrreclusive3545 Thanks, btw, are you going to upload none distilled models? that would be nice

@sweetmax797 I have the 720 T2V FP16 uploading now, says about an hour.

@mrreclusive3545I think you need to upload all the regular ones, otherwise, if someone else uploads the rest, the Civita moderators might unpublish yours

@sweetmax797 ill do my best to get them all uploaded ;) i just worry that by the time I do, new ones will be out, lol

@sweetmax797 also, just some info, as im playing with the 4step lora on the fp16 model, people are fine, but some objects, not so much, what I am doing is running the lora at 0.5 strength with 8 steps, seems to give better results on things like cars.

@mrreclusive3545 Yes, it is rank 32 and the first one in the line. There will also be rank 64. As of now, they haven't finished uploading the models. This LoRA was released at the same time we uploaded it here. Light2xv hasn't yet published any LoRA in their repository.

@sweetmax797 this is going to take awhile, first one stopped halfway through, starting over...

@sweetmax797 First one is uploaded.. finaly, lol only took 4 restarts, 720P I2V FP16 is next.

@mrreclusive3545 lol congrats! good job :)

@sweetmax797 have you tried the lightx2v lora on the 480 non distilled? i plan on testing all non distilled models against the lightx2v lora, but i dont do I2V so just curious if you tried it.

@mrreclusive3545 yes shift 5, eluer normal, 4 steps cfg 1, changing these leads to bad quality in my setup.

@sweetmax797 Thanks, I'll include that when I upload them. Also, Have you tried running no shift at all? I don't use any shift settings at all and been happy 90% of the time.

@mrreclusive3545 Yes, I tried it. I get a very noisy, cartoonish output, and below 3.5, I'm getting artifacts

@sweetmax797 strange, wonder why it works so well for me, but again, maybe its an I2V thing, I only very rarely do I2V

@mrreclusive3545 wam is less sensitive to shift value, but I'll pay more attention to it later, test it with different models maybe i get some results similar to yours

As of now, there is one lightning released for T2V about 2 hours ago. You can try that one

I love workflows that are simple "on point" like that. However, this is using much larger "all in one" HunyuanVideo-1.5-Rapid-AIO" 19gb model. I guess I would rather off load only vae and clips than everything including the model itself as I have only 12Gb vram gpu. I will make a new comment if I can get this one to work.

Yes, it works flawlessly. Thank you very much. Now I don't have to bother with official workflows. I just wish that I had better gpu than rtx 3060 12gb. I can now leave LTX 13b distilled behind me, even if HN1.5 is slower than LTX, but heck of a lot faster than WAN for me anyways.

@PirateGirl wan was a fun one for me, hated wan when it first came out, dog shit slow compared to hyv, then 2.2 came out, and it was fast as fuck.
now we get hyv1.5, slower then hyv in my setup, but we will see how it goes once more things come out. (i don't ever run any kind of cache because it distorts to much)

@Phr00t_ can i upload your aio to here https://civitai.com/models/2162731/hunyuan-video-15?modelVersionId=2435586 ? ill link to your hf if you wish.

Да, некоторые проблемы связаны с квантизацией и тем, как она реализована. Это новая версия, так что требуется время, чтобы решить ошибки и другие проблемы. Лично мне больше нравится Wan 2.2, особенно в физике и промптинге. В терминах качества необходимо проводить сравнение полной версии; FP8 чаще всего бывает удачным или неудачным. Более того, WAN 2.1 и WAN 2.2 имеют богатую экосистему LoRAs, которой у этой модели нет. WAN как SDXL, его просто нельзя игнорировать, независимо от того, насколько хороша следующая модель.

Честно - думал, что наконец-то будет конкурент Ван 2.2, но после некоторых генераций понял, что модель мертворождённая. Во первых - из-за тяжёлого Вае, и можно словить легко вылет из Комфи, а во вторых - мне не нравится качество видео в Хуньюане, а держит констинстенцию довольно так себе. Так что я предпочитаю Ван 2.2, тем более у меня свои воркфлоу есть + есть своя модель ван 2.2 10 шагов и кайф. Куда быстрее и в итоге качественнее+ комьюнити у Вана намного сильнее

yeah, im not sure how it does on low vram, but even on a 4090 its technically slower then the wan2.2 dual model setup, but for me that's only because decode. the sampling for me is actually faster, but the decode is slower, but still not as slow as the wan2.2 single model setup, decode on that takes like 4x longer then it does to sample.

and yeah, the distilled models don't do to well with liquid in my experience, the 16gb models do it fine, but its like to make it 8gb they removed that motion layer, making liquid all slow and viscose, maybe be good for another substance, but having issue's prompting translucent white fluid, lol