I absolutely love manga—do you love it as much as I do? This workflow is designed to animate manga while preserving its original aesthetic, ensuring the final result doesn’t just turn into generic "AI-slop" anime. It focuses solely on animating the existing panels, without hallucinating or inventing new, non-existent elements.

You could take a standard Minimax H3 workflow and, with some careful prompt engineering, produce a beautifully animated panel—even without using LoRAs. However, since a typical page contains 3–4 panels, animating a single 30-page chapter would require writing around 110 individual prompts. Sounds like a huge hassle, right? That’s exactly why I created this workflow: you simply upload your manga pages and get the result—no prompts, no headaches!

Now, having set your expectations high, let’s talk about the drawbacks:

1) In this first version, I prioritized preserving speech bubbles and ensuring consistent output. To prevent the AI ​​from hallucinating new events and to keep the original imagery intact, I had to limit the range of motion; as a result, some scenes might be less dynamic than you’d ideally like. 2) Occasionally, you may still encounter unsuccessful generations, so some pages might need to be re-generated.

3) You can control the liveliness of the motion in the image by adjusting the strength of the LoRA h3_fun_copy_000003500.safetensors: a value of 1.0 offers high stability but very limited movement, while 0.4 provides more movement at the cost of stability. You can also fine-tune the motion using the "video shift" parameter—a value of 12 results in stability but a somewhat stiff look, whereas 6 yields a livelier, though less stable, result.

4) There is no audio for now; achieving stable visuals was already a significant challenge. I plan to add audio generation—including scene soundscapes and character dialogue—in future versions of this workflow.

5) Videos are generated for each panel individually, so you will need to stitch them together using a video editor (I recommend FilmCraft; it’s lightweight, free, and open-source). It only takes about 10 minutes—don’t be lazy, lol! I considered integrating this directly into the workflow, but keeping previously generated segments in memory while new ones are being created would place an immense load on your PC. Furthermore, since you might occasionally need to re-generate a specific segment, a simple, automated, sequential stitch wouldn't work. If you have any ideas on how to implement this within the workflow, please share them in the comments! 6) Combat scenes or non-humanoid monsters tend to result in highly unstable animations.

7) I’ve included a custom node that automatically processes manga pages from a folder one by one (so you can download a bunch of pages, drop them into a folder, and go "touch some grass" while they’re animated sequentially). This serves as an alternative to a standard image loader. I also use a custom node to determine the correct reading order of the panels on a manga page. Additionally, I couldn't track down the original source for the bounding-box (bbox) panel detector I’m using, so I’ve uploaded it to my Google Drive—you can find the link inside the workflow.

Phew, I think that covers everything! I’d love to see the manga segments you animate—enjoy!

P.S. — Minimax h3 might be overkill for this task; I’m considering swapping it for a lighter model fine-tuned for an anime aesthetic. I already have a couple in mind, but I’d be happy to hear your suggestions!