Anima Prompt Architect

ComfyUI Anima Prompt Architect

Turn an image into an Anima prompt with PixAI Tagger v1.0 BF16 and MiniCPM-V 4.7 1B Heretic native INT8 ConvRot. Everything runs locally.

The node returns positive and negative prompts, structured analysis JSON and all 30,877 raw PixAI tag probabilities. It supports Anima Base, Aesthetic and Turbo.

Install

Clone this repository into your ComfyUI custom_nodes directory:

git clone https://github.com/NidAll/ComfyUI-Anima-Prompt-Architect.git

Install requirements.txt with ComfyUI's Python environment. A recent ComfyUI with native int8_tensorwise ConvRot operations and a compatible Comfy Kitchen CUDA backend is required. The integration has been tested with Transformers 5.15.1; the small official MiniCPM 4.7 runtime is included in the package.

Download these two files from Hugging Face:

| Model | Download | Approximate size | |---|---|---:| | PixAI BF16 | pixai-tagger-v1.0-bf16.safetensors | 0.97 GB | | MiniCPM INT8 ConvRot | minicpm-v4.7-1b-heretic-int8-convrot.safetensors | 2.66 GB |

Place both directly in:

ComfyUI/models/anima_prompt_architect/
  pixai-tagger-v1.0-bf16.safetensors
  minicpm-v4.7-1b-heretic-int8-convrot.safetensors

Each file includes its weights and required configuration, preprocessing and vocabulary/tokenizer assets. No model folders, sidecar JSON files, separate projector or additional tokenizer download are needed. Embedded assets are data, not executable code. The node never downloads models during execution.

Restart ComfyUI, import the example workflow, select an image and run Anima Prompt Architect under Anima/prompt. The positive/negative outputs connect to standard text encoders. Preview Any nodes display text and analysis. An API graph is also included. One RGB image is accepted per execution; split batches upstream.

Controls

• Variant: Base uses the recommended quality/score prefix and negative. Aesthetic omits score tags from both prompts. Turbo has an empty negative at its recommended CFG 1; the general positive-tag convention is retained.
• Detail: adjusts the tag budget. MiniCPM adds a short description of visible framing, spatial relationships, lighting and artistic details.
• Fidelity: strict uses stronger identity/style thresholds, filters defined contradictions and strips unsupported clothing modifiers. Balanced retains more visual observations.
• Confidence thresholds: control general, character/copyright and style tags. Defaults favor fewer uncertain tags; thresholds are not universal guarantees.
• Artist: explicit comma-separated artist names receive Anima's @ prefix. Artist identities are never guessed automatically from style labels.
• Safety: infer an Anima safety tag from PixAI ratings, override it or omit it.
• Slices: increase visual crops for small details at a memory/time cost. The released 16x visual compression remains intact.

Tags use lowercase and spaces, preserving underscores in score_*. Ordering is quality/meta/safety → subject count → character → series → artist → general tags, followed by natural language. These rules follow the official Anima recommendations. There is no separately documented official Turbo quality-tag recipe.

Precision and execution

PixAI's floating model weights and inference are BF16; sigmoid is FP32. The loader rejects FP32 model weights. Original aspect-preserving 1008px preprocessing is retained.

MiniCPM has 137 native INT8 ConvRot matrices. Its vision tower, visual merger, embeddings/output head, boundary blocks, normalization and sensitive recurrent controls remain BF16. FP32 row scales are quantization metadata. Native ComfyUI operations dispatch into actual Kitchen integer kernels; this is accelerated quantized execution, not an INT8 storage wrapper around dequantized matrix math.

ComfyUI patchers own sequential loading and unloading. Packed-byte gate/up and QKV/Z fusion reduce projection calls; Kitchen accelerates recurrent decoding. There is no persistent global model cache or ComfyUI core patch.

INT8 is not guaranteed to outperform equally optimized BF16 on every GPU. Prompt wording can change between precisions. See validation and limitations.

Reproduce the artifacts

Use local original model directories with their configuration/tokenizer assets:

python scripts/prepare_models.py pixai --source /path/to/pixai-original \
  --output /path/to/pixai-tagger-v1.0-bf16.safetensors
python scripts/prepare_models.py minicpm --source /path/to/minicpm-original \
  --quantizer-root /path/to/comfyui-native-quantizer --device cuda \
  --output /path/to/minicpm-v4.7-1b-heretic-int8-convrot.safetensors

The MiniCPM conversion uses ComfyUI Native Quantizer, with the included conservative recipe when the installed quantizer does not yet have the built-in family. Source checkpoints are preserved; existing outputs are never overwritten. Conversion is offline and optional.

Tests and attribution

Run portable tests with python -m unittest discover -s tests -p 'test_fusion.py' -v and python -m unittest discover -s tests -p 'test_artifact.py' -v. With ComfyUI and its dependencies installed, python -m unittest discover -s tests -v also checks loading and deployment validation. CUDA correctness tests run with python tests/test_native_cuda.py -v. scripts/benchmark.py compares compatible single-file BF16/INT8 artifacts and profiles actual native integer execution.

Code and model derivatives are Apache-2.0; see LICENSE and NOTICE for upstream sources and pinned revisions. Anima weights are not included and retain their own licensing terms.