(This model is for Anima-V1.0, the "preview2" is the version of this model not Anima)

⚠️ To use this model, you will need custom nodes and workflow. Find them inside "anima_control_tools.zip". Having an issue? Ask in comments (and they may already cover your problem, take a look at them!)

What’s new in Preview-2

The main thing is resolution. Preview-1 only ever saw 512. This one trained at 512, 768 and 1024, so it actually holds up at the bigger sizes now, and the bodies are a lot less broken up there — way fewer of those melted or split-in-half moments. 1024 is honestly where it looks best.

It also learned from a lot more art this time, roughly double the images Preview-1 had, which shows up as cleaner anatomy and steadier poses across the board. And it’s better at every resolution I support now, not just the new big ones — the higher you go, the bigger the jump.

There’s also a little ComfyUI node that gives you one more thing to try when a pose almost works but not quite. Usually you’d just keep rerolling the seed. This lets you also change how the skeleton itself is drawn: thicker lines, a colored puppet version, a heatmap, or with the hands or face dropped. The plain thin one is still what you want most of the time, since that’s the only kind it actually trained on, but every so often a different render gets a stubborn pose to land when the seed won’t. It pulls the pose off a reference photo for you too, so you’re not drawing skeletons by hand.

For the technically curious: it’s still a rank-16 control-LoRA sitting on the frozen Anima model with a zero-init skeleton embedder, so at strength 0 you get the base model back exactly, untouched. The real changes are the bigger corpus (about double Preview-1’s ~3,900 examples) and training across all three resolutions for 8 epochs instead of 512-only. On held-out poses the pose-accuracy score went up at every size, with the biggest gains at 768 and 1024.

What’s still off, honestly

• It doesn’t always follow the skeleton. Even a clean, correct one. I’ve had a clear walking pose come out as someone just standing there.

• Part of why: a thin stick figure on a black background isn’t really how anime gets drawn, so the model only half-reads it. And the ugly stuff (fused hands, general mush) shows up right where the skeleton is busiest.

• Anything dynamic — running, jumping, sitting — is still the shakiest.

• On short or lazy prompts the style drifts toward one samey default. Give it a proper prompt and it comes back.

What I’m working on right now

Preview-3 is about that “doesn’t follow the skeleton” problem, which turned out to be more interesting than just “train it harder.” The thin-lines-on-black signal seems kind of foreign to the model, so before I sink a couple days into a real training run, I ran a small bake-off on the pose signal itself: take the same poses, draw them a few different ways — the current thin lines, a thicker high-contrast one, a colored body-part “puppet,” and a depth/mannequin look — then measure which one Anima actually follows, and how much quality it costs. That’s how I’ll decide what the skeleton should even look like. The smaller fixes ride along with it: a lower default strength, dropping the noisy hand points, and more varied captions so the look stops flattening.

The other half is the detector. The one I’m using was made for photos, not anime, so on art it hands me messy skeletons, and a messy skeleton wrecks everything after it. So Preview-3 is probably going to wait until I build a pose detector meant for anime, a sort of “DWPose for anime,” which I’m starting on right after this. Preview-3 is then the actual retrain: clean detector, whichever render won the bake-off, and a lot more dynamic poses in the data. No date on it. It hangs on that detector working.

Then 1.0

If Preview-3 lands clean, that’s the one that drops the “preview” label, and it’ll train on a good deal more data again, not just the same pile pushed further.

Also available on Hugging Face: https://huggingface.co/Claquasse/Anima-Control-Pose

⚠️ The main file here is the adapter weights (adapter_model.safetensors). The ComfyUI custom node, both workflows, and the skeleton script are attached as anima_control_tools.zip, with pose_control_demo.json also attached on its own for drag-and-drop. You have an INSTALL.txt file inside of the archive to guide you. Everything is mirrored in the Hugging Face repo above.

Anima Control — Pose (Preview-1)

See the full description on the original page