Hi everyone,
I'm a solo indie developer working on a 2D side-scroller. Like many of you who have ever tried to animate a character from scratch, I hit the same wall over and over: I design a character once, but then I have to redraw it 8–24 times for each action – running, attacking, idling, jumping… It eats weeks, and honestly, it kills the momentum when you just want to prototype a mechanic.
After spending yet another weekend pixel-pushing a simple walk cycle, I thought: there has to be a better way. I know AI can generate images, but most tools either produce inconsistent frames (character morphs between poses) or require complex setups. So I decided to build something myself.
What I built: aifps.top
It's a web tool that takes a single character PNG and a motion description (like "running" or "casting a spell"), then generates a full sequence of 2D frames – ready to drop into Unity, Godot, or any engine.
How it works (from a developer's perspective)
- Upload your character art (clean, centred PNG works best).
- Pick a preset action (run, attack, idle, cast) or write a custom prompt describing the motion.
- The AI generates a frame sequence. You can adjust chroma-key settings (tolerance, smoothness, edge cleanup) to get perfect transparency.
- Export as individual transparent PNG frames or a looping GIF – drag straight into your game project.
Two generation modes: Normal for quick iteration, and High-Energy when you need more precise motion details.
The technical trade-off I made
The hardest part is temporal consistency – keeping the character's appearance stable across frames while the pose changes. Pure text-to-video approaches tend to morph the character, which is useless for game sprites. So I went with an image-driven pipeline: you provide the reference image, and the model drives the pose from your prompt. This way the output matches your existing art style.
The chroma-key step is separated as a post-process, giving you full control over the alpha channel – essential for compositing in any game engine.
Honest limitations (because we're all developers here)
- No free trial available. Running AI inference costs real money, especially the high-energy mode. As a solo developer bootstrapping this project, I can't afford to give away unlimited free generations. I've priced it as low as I can while keeping the servers running – each generation costs me compute time.
- It's a hosted web tool, not open-source. If there's enough interest, I'll happily open-source the export and chroma-key utilities.
- Quality depends on your input. Clean, centred character art works great; highly detailed backgrounds or extreme aspect ratios can confuse the model.
What I'd love from you
If you're also an indie dev or pixel artist, check out the demo video on the site to see what it can do. I'd really appreciate honest feedback on the output quality – tell me where the motion looks off, which action presets you wish existed, or how it fits into your workflow. I'll be reading every comment.
Hope this saves you some of those weeks I lost. Cheers!
Top comments (4)
This is a strong use case because the pain is concrete: repeated frames, timing, and cleanup. I would be curious how you judge the output beyond visual smoothness. For sprite tools, consistency of silhouette and editability often matter more than raw generation quality.
Great point — you're absolutely right that silhouette consistency and editability matter more than raw visual polish for a sprite tool. That's actually one of the main reasons I chose an image-driven pipeline over pure text-to-video.
For silhouette consistency, the model uses the uploaded character PNG as a structural anchor. The pose changes are driven by the prompt, but the overall shape, proportions, and key contours stay tied to the original image. I also apply a post-processing step that checks for frame-to-frame silhouette drift and flags frames where the outline shifts too much — so the user knows where to tweak.
On the editability side, the chroma-key step is deliberately kept separate from generation. Users get full control over tolerance, smoothing, edge anti-residue, and stroke parameters before exporting. This means they can clean up the alpha channel themselves rather than relying on the model's guess. The exported frames are individual transparent PNGs, not a baked composite — so they can be dropped into any engine and further adjusted.
Really interesting approach. Separating temporal consistency from background removal seems like a smart design choice for game assets. I'm curious—have you tested how well the generated animations stay consistent across longer sequences like combo attacks or multi-stage boss animations?
Continuous animations can be split into multiple segments and combined together — as long as the reference image remains the same, character consistency is perfectly guaranteed.