Thinking Nodes: Let the LLM Write the Prompts in Your Anime Workflow
The wall in AI anime production is not image quality, it is serial work: faces drift, assets get lost between tools. A Japanese creator shows how PixAI Studio solves it by putting LLM nodes inside the pipeline so the model writes its own edit prompts.
AI-assisted draft. Reviewed and edited by the Phosphene team before publication.
Single-image generation is solved. Any decent model can hand you a character that would have made a splash two years ago. The problem that actually kills AI animation projects comes one step later: make the second frame, and the face shifts. Make a third, and the jacket is a different jacket. By the time you try to animate, you are not animating your character anymore, you are animating a stranger.
Most creators hit this wall and reach for a better model. The more useful fix is structural: stop thinking in single generations and start thinking in workflows, where the pipeline itself keeps the character intact and the LLM writes the prompts for you.
The tool-hopping tax
Here is the workflow most people actually run, whether they admit it or not:
- Generate the illustration in one tool.
- Open a retouching app to fix details.
- Re-upload the result to a video generator.
- Lose the source file somewhere in a downloads folder.
Every handoff between tools is a place where consistency dies. The face changes because a new model reinterprets the reference. The outfit changes because the edit tool never saw the original prompt. And the asset management fails because nothing is connected to anything else. The Japanese AI creator Ai-Hakase calls this the "tool ladder," and it is the real reason anime projects stall: not model quality, but fragmentation.
The opposite of the ladder is a workflow where generation, editing, video, and asset storage happen on one canvas, and data flows from node to node without a single manual export. That is the design idea behind PixAI Studio, a browser-based node workspace, and the workflow Ai-Hakase built on top of it is worth studying even if you never open the tool.
Let the model write the prompts
The interesting part is not the canvas. Node pipelines exist elsewhere, most famously in ComfyUI. What makes this workflow different is the thinking layer: LLM and VLM nodes that do not just transform an image, they write the instructions for the next step.
You drop in a source image. A VLM node looks at it and produces the image-edit prompt as structured JSON, ready for the edit node. You want a gacha game screen instead of a plain portrait? The VLM knows what a gacha screen is, and it writes the edit prompt for you. You want a dating-sim dialogue UI for the second frame? Same thing. The prompt for each frame is generated, not hand-written.
The practical consequence is that a beginner and a professional start from the same place. The knowledge gap is absorbed by the model. The human's job changes from "write a perfect prompt" to "review a generated prompt and say what is wrong."
That review step is not decoration. The workflow deliberately runs the VLM at a temperature of 1.0 with a generous token budget, which trades consistency for flexibility: the model produces more varied, more human-sounding instructions, and occasionally garbles the output. The documented fix is to open the node and edit the text directly. This is the healthy version of automation, the model drafts, you approve.
The one-image pipeline
The full workflow takes a single source image and produces three frames plus a voiced animation. The source can be an upload, or a fresh generation from PixAI's Tsubaki.2 model via a prompt helper node, which means you can start from a plain-language description in Japanese or English and never touch an image editor.
The frames are the fun part. Frame one is a gacha pull screen, complete with the game UI language that model has learned. Frame two is a dialogue screen in visual-novel style. Frame three is a close-up romantic scene. Each frame is produced by the same pattern: the VLM node writes the edit prompt, you can append personal requests in plain language (make it 16:9, brighten the background), and the image edit node renders the result.
Then the video stage takes over, with two manual steps. A prompt creator group assembles the video-generation prompt from the frames, and you copy that generated prompt into the video node and enable Omni reference before running it: multiple assets, images and audio, passed to the model in one call. The output is a voiced animation built from those three frames. The prompt handoff and the reference toggle stay manual, so the video stage is assisted rather than fully automated.
For the impatient, the workflow supports group execution: instead of running nodes one by one, you run a group and every node inside fires in sequence. The whole asset set can be produced with a few button presses.
Consistency as a system
The character consistency trick is worth stealing even if you never use PixAI Studio. It comes down to anchors and compounding assets.
An anchor is a reference image that stays connected to the pipeline. When you want a costume change or a new angle, you do not rewrite the prompt from scratch and hope for the best. You keep the anchor image wired into the reference node, so the identity stays fixed while the composition changes.
The compounding part is subtler. PixAI Studio lets you import your past favorites and generation history directly into the canvas. Every finished work becomes source material for the next one. Your production library grows as you work, instead of scattering across folders. Ai-Hakase describes reusing old images as fresh material for new videos, which sounds obvious until you realize most creators never build the habit because their tooling makes it annoying.
Templates versus the blank canvas
PixAI Studio offers two starting points, and the choice maps to a real learning path. Beginners should clone a template and swap the inputs: the node connections are already made, so you learn how the pipeline works by changing the image and the personal requests. Once the structure makes sense, you build your own workspace from an empty canvas and design a custom pipeline with your own thinking nodes and edit nodes.
This is the same progression that made ComfyUI workflows popular: first you run someone else's graph, then you understand it, then you build your own. The difference is that in a managed browser workspace the setup ritual is gone, no custom nodes to install, no model files to place, no environment to break.
What transfers, what does not
If you take one idea from this workflow, take this: in serial AI production, the prompt is a product of the pipeline, not an input you sweat over at the start. A model can write edit prompts for you, check them, and iterate, as long as you build the structure where that can happen. That insight works in ComfyUI, in any node tool, and in a template-first product like ours, where the whole point is that a well-designed starting point beats a blank page.
Two caveats from the original writeup are worth remembering. The VLM node consumes credits on every run, roughly 400 per processing pass in this workflow, so automation is not free and you will want to review prompts before spending on the expensive steps. And the temperature trick cuts both ways: flexible thinking means occasional garbage output, which is fine only if you keep a human in the loop to catch it.
The takeaway is not that PixAI Studio is the only place to do this. It is that the tool ladder is optional. Once you treat an anime production as one connected pipeline, with an LLM writing the intermediate prompts and anchor images holding the character together, the bottleneck stops being consistency and becomes the actual creative choices.