NidAll/Anima-Lightning-INT8-ConvRot

Anima-Lightning INT8 ConvRot

Native ComfyUI INT8 ConvRot quantization of aina-tech/Anima-Lightning, the four-step distilled Anima text-to-image model. This release preserves the source model's architecture and distilled sampling trajectory. No additional training or distillation was performed.

The checkpoint contains the diffusion transformer and Anima LLM adapter in one native ComfyUI safetensors file. Use it with the usual Anima Qwen3 text encoder and Qwen Image VAE, which are downloaded separately.

Download and load

Download Anima-Lightning-int8-convrot.safetensors and place it in ComfyUI/models/diffusion_models/. Select it in Load Diffusion Model in an Anima workflow. Use a recent ComfyUI and comfy-kitchen version with int8_tensorwise ConvRot support. This file uses native ComfyUI tensor names and quantization metadata; the upstream Diffusers loading example does not apply to this artifact.

Required sampling settings

Connect Load Diffusion Model → ModelSamplingAuraFlow → KSampler.

Setting	Value
ModelSamplingAuraFlow shift	1.0
ModelSamplingAuraFlow sampling	flow
Sampler	euler
Scheduler	simple
Steps	4
CFG	1.0
Denoise	1.0
Negative prompt	Empty; CFG 1 uses conditional inference
Recommended starting resolution	1024 × 1024

These settings produce [1.0, 0.75, 0.5, 0.25, 0.0]. Native Anima defaults to shift 3.0, so apply the shift override. For a custom sampler graph, use ManualSigmas with that complete list, KSamplerSelect = euler, BasicGuider and RandomNoise.

Keep the distilled four-step schedule. The upstream full-step Anima schedule and higher CFG settings are not the intended runtime for this checkpoint. See the original model card and distilled_generation_config.json for the source runtime contract.

Quantization

Property	Value
Weight format	int8_tensorwise with convrot: true
Rotation	Regular Hadamard ConvRot, group size 256
Quantized layers	280 transformer attention and MLP linears
Scaling	FP32 scales per output row
Preserved precision	BF16 inputs, outputs, embeddings, normalization, timestep/AdaLN modulation and the full LLM adapter
Checkpoint size	2.424 GB / 2.257 GiB
Source revision	ad1cb61e3d569dc01b748f83ceb58a8199583cee

The published source components are transformer/diffusion_pytorch_model.safetensors and text_conditioner/diffusion_pytorch_model.safetensors. The Diffusers transformer keys were mapped back to the native Cosmos Predict2/Anima layout; conditioner keys are stored under llm_adapter..

Converted with ComfyUI Native Quantizer, version 2.2.0, using its int8-convrot format.

Validation and limitations

The checkpoint passed the converter's payload, metadata, tensor-shape and scale checks. All 280 candidate linears passed its weight reconstruction gates. A four-step ComfyUI generation was reported working by the publisher.

INT8 quantization introduces numerical differences, so generated images can differ from the source BF16 model. No standardized image-quality or performance benchmark is provided. The original model's limitations, including text rendering and complex anatomy, still apply. The model is intended for anime, illustration and stylized artwork.

License and attribution

Distributed under the CircleStone Labs Non-Commercial License v1.2. Rights to use the model and this derivative are granted directly by CircleStone Labs LLC under that license. Applicable upstream license terms remain in force. See NOTICE.md for the required attribution and modification notice.

This is a community quantization of the aina-tech release. It is not an official or endorsed release from aina-tech, CircleStone Labs, Comfy Org or NVIDIA.

Upstream lineage: Anima-Lightning → Anima → Cosmos Predict2.