H3-Image keeps my characters on model from one reference sheet. Is training a character LoRA still worth the evening?

by @lio_toons in c/character-animation ·

Since MiniMax put the H3-Image weights out this week I've been feeding it my little 3D mascots (the toaster repairman, the night-shift owl) as a single turnaround sheet and asking for new scenes. Out of 40 tries, maybe 33 were on model: same proportions, same scuffs on the overalls, same eye shape. That's better than the LoRA I spent two evenings training back in the spring.

Where it still slips for me:

Tiny details like a logo on a cap drift after a few edits in a row.
Two characters in one frame - the second one starts borrowing features from the first.
Extreme angles (top-down, worm's eye) are where it guesses.

So my question for people who do this every day: in what situations do you still train a LoRA? I can think of "a style the base model has never seen" and "batch jobs of hundreds of frames", but I'd like to hear real cases before I delete my training folder.

#minimax #h3-image #lora #character-consistency

158 points · 20 comments

Comments

  • @irischopra_art935 · · 15 points

    As a 3D person: if you already have the models in Blender, render a quick clay turnaround at 8 angles and use that as the reference. The extreme-angle guessing mostly goes away because you've literally shown it the top of the head.

  • @lio_toons · · 8 points

    I do have the Blender files! Never thought of feeding it renders instead of my polished sheet. Trying tonight.

  • @meinascimento_studio · · 13 points

    I remember when keeping one face for three shots in a row felt like a miracle. Now we're arguing about logo drift on a cap. What a time.

  • @natashaweber_ai220 · · 17 points

    Batch work is still LoRA territory for me. When a client needs 300 product-mascot frames, I can't babysit every reference edit. A LoRA gives me a predictable failure rate I can plan around.

  • @pavelsilva_digital · · 9 points

    "Predictable failure rate you can plan around" is the most professional sentence in this thread.

  • @katyaitomedia512 · · 8 points

    Beginner question: is H3-Image the same thing as the H3 video model, or a separate download?

  • @dmitrischneider_motion784 · · 15 points

    Separate weights, same family. The image model was built on top of the video model, which is probably why it's good at "same character, different moment".

  • @lucaswatanabe_motion448 · · 8 points

    Two characters in one frame is the classic problem. What helps me: give each character its own reference image instead of one combined sheet, and describe them by role in the prompt ("the tall one holding the lamp") instead of by name. Names seem to blur together.

  • @dmitrischneider_motion784 · · 13 points

    Can confirm the separate-images trick. Also keep the two references at the same resolution, otherwise the bigger one wins.

  • @lio_toons · · 3 points

    Separate reference images, noted. I had them both on one sheet to save time, which was probably the whole problem.

  • @evelyngarcia_ai47 · · 6 points

    The logo drift thing: I paste the logo back in compositing at the end. Takes 30 seconds per frame and saves an hour of rerolling.

  • @lio_toons · · 2 points

    Honestly that's what I'll end up doing. The cap is 40 pixels wide, nobody needs it generated.

  • @noor_anim · · 4 points

    For 2D sheets I still train. Reference editing nails the face but it keeps "fixing" my line weight toward its own idea of clean anime lines. A LoRA is the only thing that keeps my wobbly ink style.

  • @lio_toons · · 9 points

    That's a good one. Style the model wants to correct is basically the "style it hasn't seen" case wearing a different hat.

  • @adrianmillermedia · · 4 points

    For anime series work I'd still keep a LoRA as the "source of truth" and use reference editing for speed on top. Two tools, not a replacement.