Anima ONE STUDIO (TJ)

Anima ONE STUDIO (TJ) is the newest node in the TJ NODE STUDIO ONE pack, and it's the one I'd point a beginner at first. It wraps Anima - ComfyUI's own native anime/illustration model, roughly 2B parameters and proudly non-photorealistic - into a single node that does text-to-image, inpainting, and two image-guided modes, plus a Turbo toggle, without you wiring a single connection. In a pack whose other nodes drag in Impact Pack, KJNodes, and friends, this one is refreshingly self-contained: the base pipeline needs no extra custom node packs at all, just model files.

That's the whole pitch of the ONE STUDIO family, honestly. These are app-in-a-node UIs, a direct answer to the "why does every workflow I download need ten missing packs" complaint that haunts the ecosystem. You drop the node on the canvas, pick a mode, hit the button inside the node. No graph to read.

How it works

Anima is native to ComfyUI - architecture lives in comfy/ldm/anima, and the official templates ship in ComfyUI itself. The node reproduces those templates in its own UI. Two things are worth understanding because they explain the settings you'll see:

Turbo is not a separate model. Anima Base 1.0 samples at 30 steps / CFG 4. Flick the TURBO switch and it loads the anima-turbo-lora-v0.2.safetensors LoRA on top of the same Base 1.0 checkpoint and drops sampling to 8 steps / CFG 1 - the classic turbo-LoRA trade, much faster, marginally less polish. There's also a separate older anima-preview3-base checkpoint, but that's a T2I-only curiosity; ignore it unless you're comparing.

The guided modes aren't ControlNet. Inpainting, Any Control to Image, and Depth Control to Image use LLLite model patches (a light LoRA-style injection into attention layers - the modern way to do structure control on DiT models). Inpainting takes your mask and re-draws inside it; Any Control guides from an image with an optional mask; Depth Control first runs Depth-Anything-3 to build a depth map, then guides from that. So the thing people think is ControlNet is really a lighter-weight cousin, and it works fine for what it's meant to do - no preprocessor chains to babysit.

The inputs that matter

The node has no required inputs. The two optional ones are both advanced:

• prompt_override - a multi-line string prepended in front of the internal prompt. Handy if you want to force a style across generations from outside.
• pipe - the PromptDB pipe from the author's separate ComfyUI-TJ_NODE pack. At generation it overrides prompt, seed, steps, CFG, sampler and scheduler with whatever PromptDBLoader(TJ) had selected; the model stays whatever you set in the node, and your UI never changes. If TJ_NODE isn't installed, the socket just sits inactive.

There's one output: image, the most recent image the node generated. Note what that means: this node runs off its own ▶ Generate button, not the top-of-ComfyUI RUN. Hitting RUN merely re-emits the last result from the output slot - a genuinely common "why is nothing happening" moment. The ⚠ in the node's own help text says exactly this.

Everything else lives in the ⚙ Settings panel: pick the text encoder, VAE and diffusion model, toggle languages, set a save subfolder.

Install and models

Install the pack either way:

cd ComfyUI/custom_nodes
git clone https://github.com/designloves2/ComfyUI-TJ_NODE_STUDIO_ONE.git

or use ComfyUI Manager and search for ComfyUI-TJ_NODE_STUDIO_ONE, then restart. Then grab the Anima files - text encoder qwen_3_06b_base.safetensors to models/text_encoders/, qwen_image_vae.safetensors to models/vae/, anima-base-v1.0.safetensors to models/diffusion_models/, and the Turbo LoRA to models/loras/ if you want it. The three LLLite patches (anima-lllite-*-v2.safetensors) go in models/model_patches/, and Depth mode needs depth_anything_3_mono_large.safetensors in models/geometry_estimation/. All of it is open-weight on Hugging Face. When files don't show up in the dropdown, hit ↻ Refresh Models before you restart anything.

Where people get stuck

The Generate-vs-RUN thing is the big one. After that: inpainting masks are drawn in the inline painter, so if your masked area looks wrong, check the Invert toggle before re-rolling. And the community is still feeling Anima out - threads about inconsistent seeds and hair prompting are common, because the model is young, not because you broke it. Turbo won't engage until the LoRA is actually downloaded. All in all, if you want to see what this "one node is a whole app" idea feels like without installing five packs to get there, this is the node to try first.