all / g / s / fur / art / ai / 3dcg / f / rp / req / p2p / dis / lit / kaki chat
[Settings] [Home] [Catalog] [Search] [Thread list] [Stats] [Reports] [Watcher] [PMs] [Admin]
[Return] [Bottom]

Posting mode: Reply
BBCode
(for deletion)
  • Allowed file types are: gif, jpg, jpeg, png, webp, webm, mp4
  • Maximum file size allowed is 25000 KB.
  • Images greater than 255 * 255 pixels will be thumbnailed.
  • 201 unique users in the last 30 minutes (including lurkers)
    First
    [1]
    		Last
    File: 192951_00001_audio.webm
    (10 MB, 1408x1088) ▶
    10486702
    No.29162 [Edit] [Report] AS
    ▶
    Kinda relapsed into doing this sort of thing - seeing what MiniMax H3 could do brought me back. What amazed me about this model was just how easily it could create some pretty wild stuff even without LoRAs. Using a workflow I found off Civit and the correct uncensored/abliterated Qwen VL text encoder, I could get it generating stuff like *this* out of the box on consumer hardware. Color me impressed and concerned in equal measure...

    I've even tested it with making hearts burst, and it's serviceable - just need to prompt it correctly. Speaking of which, this vid was done in ComfyUI, so it has all the prompt and workflow metadata built right in!
    >>
    No.29163 [Edit] [Report]
    ▶
    File: MiniMax_H3_48FPS_00002-audio.mp4
    (19.79 MB, 2560x1440) ▶
    20751935
    This one was done in an Uncensored MiniMax H3 Runpod instance...weirdly enough, given the prompt I was trying to get something *less* realistic, not *more* realistic! It's meant to be 3DCG, but I think it went a bit too far on the realistic end of it. Did a damn good job with the heart though, but with a bit better prompting I could probably make the damn thing wobble more. Anyway, enjoy this lady who seems to have been imprisoned pregnant...
    >>
    No.29175 [Edit] [Report]
    ▶
    Can you tell me how your getting such clean renders are you doing 1080p? what are your specs?
    >>
    No.29180 [Edit] [Report]
    ▶
    File: 200450_00001_audio.webm
    (15.77 MB, 1728x1280) ▶
    16537918
    In the first case, I generated it locally, with a 4070 Ti Super with 16 GB VRAM and 64 GB DDR5 RAM; I use ComfyUI and DaSiWa's H3 workflow (here: https://civitai.red/models/2831978/dasiwa-minimax-h3-workflows-or-t2va-or-fl2va-or-ref2va )

    The cleanness of the first render is because while I only genned at about 480P/0.36 MP, I upscaled it using the 2X-AnimeSharpV3 upscale model. Seems with the right settings it does the job well enough.

    The second one was also upscaled, but it started from a render size closer to SD on a Runpod Instance, so I wouldn't call that a particularly useful benchmark.

    Here's another gen I had the computer spit out, kind of along the same lines as the first but more intimate and, well, self-inflicted I guess. As before, you can see the full workflow for it if you drag the video into ComfyUI.
    >>
    No.29226 [Edit] [Report]
    ▶
    Well, this is annoying...the Runpod template I was using for a number of these gens has been failing to load ComfyUI due to not being able to fetch certain components, at least if I understand the error logs correctly. Not been able to find another Runpod template for MiniMax H3 that meets my needs, and trying to set one up manually is a bit beyond my skill level given how long it took me to set up my own local ComfyUI. I'll keep an eye out for templates, but with the cost of running a Runpod instance I'll probably stick with local genning at a lower quality for the time being. Apologies.
    [Top]
    Delete post: []
    First
    [1]
    		Last