[search help]

Advanced Options

Post relationships

Layout Settings

Blacklisted
    (help)
    Director
    Franchises
    Characters
    Species
    General
    Meta

    Information

    ID
    182568
    MD5
    275a668e3b1594d7e4cbc39b342fbaaa
    Size
    1920x1080 (95.2 MB)
    Type
    MP4
    Status
    Active
    Rating
    Questionable
    Score
    53
    Faves
    44 Show

    Related

    Viewing sample resized to 78% of original (view original) Loading...
    Fullscreen
    Description

    "I need a render farm for my bunnies"

    Mello's Zootopia AI Demo Reel - The Complete Saga (Minimax H3 Final Tests)

    Songs Used:
    eliderp - iridium (Knoir flip)
    Kendrick Lamar - count me out (Surphase Flip)
    Moore Kismet - I'M YOUR FAVORITE!
    Yasushi Akutagawa -メインタイトル - アルバム収録音源

    Alright now its zootopia's turn. Tried to finish this before classes started but decided to put a bit more effort in. Now its turned into a four-part thing and stuff lol. More detail on each part bellow.

    Part 1 (Render Dates: 8-22-26 - 8-23-26): These are the initial tests generated after I deemed the lora training 'complete enough.' About the first 40 or so renders were just testing the dataset and got mixed results. It was when I added more flamboyant language and explanatory prompts when things started to look a lot better. Combine that with the new starlight upscaler and fur details are massively improved compared to wan along with the fact multi-shots are possible. This was also a test to combine various loras, results are iffy as combining more loras together just seems to degrade the video or hurt character's resemblance, unlike in wan. There is no story here unlike the other parts, just your usual collection of tests.

    Part 2 (Render Dates: 8-23-26 - 8-27-26, easily my most favorite part): This one took quite a bit of time. Originally was just going to be a super short of a desert town that went into a building where Nick was, otherwise fit into the first part of the showcase. I tested multiple environments and actually found the snow to be very impressive and thus the first 20 seconds were born. That last shot in that sequence compelled me to actually turn it into a longer story that went into the night and have some conflict. One fun revision I originally planned was instead for the female fox to approach Nick while he was still cowered down at the base of the mountain, at which point he would turn, she would reveal herself, and she would extend her arm to help him up. I changed it to its current form when I accidently typed "Nick looks up" instead of "Nick looks out," thus leading him to look up at the camera during the end of the firefight shot, I thought it was a very cute moment, so I changed that bit, I think it also fits the fast passed nature of the short better. I made this before I was aware of the 'infinite' video stuff that was out there so some cuts are a little more obvious but ah well, I still like it (especially the nerdy fox).

    Part 3 (Render Date 8-27-26): A 4-hour generation spree. I knew I originally wanted to do this part after seeing that guy comment more than a handful of times (I deleted most of them). Message to that guy: your goon material has arrived, pls stop.

    Part 4 (Render Date 8-27-26 - 8-28-26): Funnily enough, I was just testing longer form prompts and generation, upwards of 15 seconds just to see what it would do. I was also testing some speech stuff and varying aspect ratios. Thats why the first third of this part is a little jumbled and disconjointed. It was only after making those initial renders that I went in on making a story that had scenes consisting of 'ups and downs.' I only got the trailer idea at the very end, which is why the more rapid cuts and stuff come later on. This was also a fun exercise in making an aesthetic for making things sound and look 'old' / 'film reel-y'. (note: this is just a joke trailer, I aint making an actual film or smth).

    Timestamps:
    0:00 Part 1: Judy
    1:13 Part 2: Nick
    2:55 Part 3: Nick & Judy
    4:04 Part 4: Movie Concept

    Link to uncompressed versions:
    https://www.melloai.art/renders/zootopia/minimax_h3/Minimax_H3_Final_Tests/

    Lora details:

    Name: "zoo_V6.1_14000":
    - Trained on 258 clips from "a certain movie" (Ranging from 5-224 frames), 14,000 steps, ~25 epochs.
    - 8 retrains later finally got something that wasn't completely buns.
    - Can quite literally do anything: multi-shots, different character color combinations with different aesthetics, different species, fast or intimate movements.
    - First time training with audio, Judy's voice is quite good but Nick and Pawbert sound a bit more rough in comparison, either generating a "close-enough" voice or a completely voice. (Was quite a pain to tag sounds and stuff)
    - All gens were done at 512p, it can do 804p gens but motion seems to be quite a bit slower, although detail is quite impressive, will experiment more

    Name: "minimax_gopro_V1_5000":
    - Trained on 42 clips of a guy wondering around with a gopro-like camera and shoving it in people's faces.
    - Initially trained for wan but it seems to have transferred over quite well, camera movements seem to be a little more restricted than wan but it gets the aesthetic across nonetheless.

    Name: "minimax_space_v1.1_7000":
    - Trained on 49 clips of various distortions, VFX, and some chaotic stuff I made in blender
    - Another dataset transfer directly from wan but also seems to have turned out quite well. Seems to have a higher failure rate than its wan counterpart but when it hits, it hits (this is most likely down to the destructive language I used when training the dataset).

    Blacklisted
      (help)

      There are no visible comments.