Anima P3 Turbo — dequantized BF16 weights for pictor / MLX
Personal backup of the four BF16 safetensors files converted from
n-Arno/Anima-P3-Turbo-AIO-Q4_K,
revision ddb1225449828d4e5d69208c008b53f838e626c5.
The original AIO checkpoint contains Anima Preview 3 with a merged Turbo LoRA,
a text encoder and a VAE. Credit for the base model belongs to CircleStone Labs,
with the Turbo/AIO and GGUF preparation credited through the source model card.
Conversion
- Source:
Anima-P3-Turbo-AIO-Q4_K.gguf. - Source SHA-256:
3290dc9abad9cf98cf1b39b491a464bd4e7bba200ed508dcf0a7e9cd8fdb9168. - Q4_K tensors were dequantized using the pinned sd.cpp/ggml implementation, then rounded to BF16 (round-to-nearest-even); retained BF16 values are preserved.
- Tensor names/layouts were mapped for the native pictor Anima MLX backend, following the xocialize/anima-mlx architecture mapping. These are P3 Turbo weights, not the public Anima v1 checkpoint.
- Only the VAE decoder is exported; the VAE encoder is omitted.
- This is a dequantized export of an already quantized model. It does not recover the original pre-quantization BF16 weights or lost precision.
conversion.jsonrecords source provenance, converter identity, tensor counts, output sizes and SHA-256 hashes.SHA256SUMScovers all uploaded payload files.
Files
| File | Component |
|---|---|
transformer-bf16.safetensors | Diffusion transformer |
text_encoder-bf16.safetensors | Text encoder |
llm_adapter-bf16.safetensors | LLM adapter |
vae-bf16.safetensors | Single-frame Wan decoder |
Weights total 5,520,953,358 bytes (5.14 GiB). All four files are needed. This repository contains weights for pictor's native loader, which supplies the architecture and tokenizers; it is not a standalone Diffusers pipeline.
Use with pictor
hf download nemurubaka/anima-p3-mlx-bf16 --local-dir models/anima-p3-mlx-bf16
zig-out/bin/pictor anima --backend mlx --model models/anima-p3-mlx-bf16 \
--prompt "anime landscape, a cottage by a lake" \
--width 512 --height 768 --steps 8 --cfg-scale 1 --cache none \
--seed 666 --output image.png
The native backend uses ER-SDE / SmoothStep sampling. Validated P3 presets use
3 / 8 / 16 steps and CFG 1. Runtime code and model preparation are maintained in
pictor; the converter version used here is recorded in conversion.json.
License and attribution
The source repository declares the CircleStone Labs Non-Commercial License.
The upstream license is included as LICENSE.md, with its source revision
recorded in conversion.json. This conversion adds no new grant of rights to
the model weights. See the source model card for the merged Turbo LoRA provenance.