---
license: apache-2.0
pipeline_tag: image-to-text
library_name: comfyui
base_model:
- pixai-labs/pixai-tagger-v1.0
- saidutta69/MiniCPM-V-4.7-1B-heretic
tags:
- comfyui
- safetensors
- image-captioning
- image-tagging
- int8
- convrot
---
# Anima Prompt Architect models

Two self-contained safetensors models for
[ComfyUI-Anima-Prompt-Architect](https://github.com/NidAll/ComfyUI-Anima-Prompt-Architect).

| File | Precision | Bytes |
|---|---|---:|
| [pixai-tagger-v1.0-bf16.safetensors](pixai-tagger-v1.0-bf16.safetensors) | BF16 | 972,940,675 |
| [minicpm-v4.7-1b-heretic-int8-convrot.safetensors](minicpm-v4.7-1b-heretic-int8-convrot.safetensors) | INT8 ConvRot + BF16 | 2,656,884,950 |

## Installation

Install the custom node, then place both model files directly under
`ComfyUI/models/anima_prompt_architect/` and restart ComfyUI. Select them in
**Anima Prompt Architect**. Each artifact embeds its required configuration,
preprocessing and vocabulary/tokenizer data. No sidecars, separate vision
projector, remote executable code or automatic model downloads are needed.

These are custom-node safetensors artifacts. Generic Transformers `from_pretrained`
and stock ComfyUI diffusion/text-encoder loaders are not their consumers.

## Precision and validation

PixAI v1.0: all 398 model tensors equal original values rounded to BF16. The
loader requires BF16 floating model weights; inference uses BF16 and sigmoid FP32.
Original 1008px aspect-preserving preprocessing is retained.

MiniCPM-V 4.7 1B Heretic: 137 native INT8 ConvRot matrices and 643 unchanged BF16
model tensors. Vision, merger, embeddings/head, boundary blocks, norms and sensitive
recurrent controls remain BF16. FP32 row scales are quantization metadata.
Native ComfyUI operations and Kitchen kernels execute the packed integer weights.
A complete generation succeeded with INT8 weight dequantization forbidden.

Packaging preserves every weight tensor exactly and verifies embedded-asset
integrity. The node has been exercised through the ComfyUI API. Precision changes
can alter descriptions; neither hallucination-free output nor universal INT8
speedup is guaranteed. Character recognition, multi-character binding, OCR and
Anima reconstruction quality have not received a broad labeled evaluation.

## Sources and license

- [PixAI Tagger v1.0](https://huggingface.co/pixai-labs/pixai-tagger-v1.0), revision
  `9fe10addf9326e292da8a85a98ea74cd91b41771`, by PixAI Labs, Apache-2.0.
- [MiniCPM-V 4.7 1B Heretic](https://huggingface.co/saidutta69/MiniCPM-V-4.7-1B-heretic), revision
  `f2ec4781396394599a524bd396133ed117b38da1`, derived from OpenBMB MiniCPM-V 4.7 1B,
  Apache-2.0. The Heretic derivative modifies the original model's refusal behavior.
- Conversion uses [ComfyUI Native Quantizer](https://github.com/NidAll/comfyui-native-quantizer).
  The custom node includes the conservative conversion recipe and reproduction scripts.

Both converted models retain Apache-2.0 licensing. See [LICENSE](LICENSE),
[checksums](SHA256SUMS), and the source repositories for original attribution.
Anima weights are not bundled and retain their own licensing terms.