Thread Rules:>Only videos made in Local AI video generators are allowed such as WAN or LTX or others>No videos made in Popvid or Grok or any other similar low-quality generators>No AI Celebs >No Prominent AI glitches for no reason>No BlackedPrevious: >>31243192
>grab still from cutscene in game>use it in WAN to generate video>do the same for genned video>paste them together>end result 10x better than arsetard's original scene
>>31255469what model is this? and is it free like minimax ?
>>31255386more things like this
Just getting started with Minimax H3 and got directed to this thread. It seems like it doesn't do uncensored content well unless it's in the initial inputs? Does anyone have any advice on getting the best results from it or the best LORAs to use for explicit gens?
>>31255622not bad for your first h3 days. i might be mistaken, but there still isn't any aio nsfw loras for t2v-ref2v. you have to use the available nsfw loras and test them out
For those wondering... Yes, you can run local generation on a 8gb rtx gpu if you use quantized models and sage attention.The results are not top quality and take much longer, obviously. But you can still generate and learn.
trying out h3
>>31255987Impressive! The movement and background moving worked well, and the bed reacting turned out surprisingly well. Was that using a reference video or something? Can I ask about the settings / prompt you used?
>>31255469Whats the best Minimax H3 Turbo Lora right now because theres so many of it out in the last 3 weeks >Whats the best Minimax H3 Turbo Lora right now because theres so many of it out in the last 3 weeks Whats the best Minimax H3 Turbo Lora right now because theres so many of it out in the last 3 weeks >Whats the best Minimax H3 Turbo Lora right now because theres so many of it out in the last 3 weeks
>>31256025I think a lot of it is just H3 doing well. I'm not using the fast version, as in my experience things like that tend to degrade quality. So this is just the normal H3 workflow running locally, gen took ~800s. I had a reference image as input, but nothing else (nor any LORAs).Compared to LTX the frame-to-frame and subject consistency is WAY better out of the box. It even keeps her mask despite a lot of the video being her turned around, and the only mistake really is it maybe interpreting the mask as a sleep mask at the end.in my experience, the best thing to guide things is just good, high quality input.
>>31256139Thanks for the info!>in my experience, the best thing to guide things is just good, high quality input.What inputs are you using? First frame / last frame, reference video, audio, multiple sources?>gen took ~800s.Impressive for that quality! What kind of GPU do you have? I've got a 3090 but gens are taking ~25 minutes with these settings:10 steps (with a turbo lora)1.5 megapixels10 second videoObviously faster would be better, but I don't want to sacrifice quality.>>31256085The Turbo LORA that came with workflow I downloaded is called: minimax_h3_fl2v_turbo_silver_dareties_comfy_pruned_v1 . I'm using https://civitai.red/models/2881362/minimax-h3-seed-hunter-latent-upscaler-seamless-video-continuation-speedups-ref-mods
>>31256155>What inputs are you using?literally just starting frame. I'm only dipping my toes in here so not doing full FLF or anything more yet>gpuI'm on a 4090. that video is 10s, 0.8 MP, 10 step, with the turbo lora (just checked).
>>31256139You should use Spectrum though.
>>31255481This came out great Show a few more these, even the little story line was fun. btw is Hogwarts Legacy inspired ?
>>31256231what is that?
>>31256258https://github.com/xmarre/ComfyUI-Spectrum-MiniMax-H3It's designed to be used with higher step counts.
>>31256265NTA. It also works on 8 step workflows.Regardless of what people say you should and shouldn't, I use an 8 step lora, comfy kitchen, sage attention, spectrum, chunk and feed forwarding, and a 2 stage setup (5 steps at 1/2 res > latent upscale > 3 steps at full res) all in one workflow. I have tested and proven each piece of it to have a benefit when combined, although in theory some of those things shouldn't play together properly.With this monstrosity of a pipeline, I get 15s of 960x1280 in 400s from my 5080, or about 800s from my 5060ti.
>>31250853>>31251605prevs>>31250923>>31252784anons>>31256085larryvrh/MiniMax-H3-Turbo-Lora
>>31256155>https://civitai.red/models/2881362/minimax-h3-seed-hunter-latent-upscaler-seamless-video-continuation-speedups-ref-modsHave you try the video continuation ?
>>31256331Most boring shit i ever seen. Please stop.
