AbstractPhil/anima-prelim-1k-r64

Current Train Effect

The model did learn the 1000 images effectively. The lora is a success, the model learned JSON and the utilization of that json to produce symbolically similar contextualizations to the subject.

To Use

Load the primary Anima model and load this lora epoch 20.

The comfy-qwen-json nodes are present here, so if you wish to use the json converter you can use it directly in comfyui with those node files. Place the files in a directory within the custom_nodes section of comfyui and install the requirements via comfyui methodology.

You are NOT required to use the json, but it does help.

The json model does not accept (tags:1.5) tagging of this nature currently.

String combine the json output with the raw tags for the best effect for now. Tags after.

Anima preliminary LoRA (rank 64, before_after subject buckets)

Preliminary LoRA finetune of CircleStone Anima (2B DiT, NVIDIA Cosmos-Predict2-2B backbone), trained with diffusion-pipe via the geolip_anima_trainer package on a ~1000-image slice of AbstractPhil/diffusion-pretrain-set-ft1 (qwen_90k).

Methodology: task_1 JSON captions trained verbatim in semantic subject buckets. caption_mode=before_after — a full VLM phase (runs/vlm/...) then a full animetimm phase (runs/animetimm/...) resuming the VLM adapter. Sparse subjects grouped by semantic similarity, oversized buckets split by attribute, weighted with a diminishing-returns num_repeats (capped ~8×); distinct human subgroups kept separate. llm_adapter_lr = 0 (adapter frozen).

NON-COMMERCIAL. Derivative of Anima — inherits CircleStone's non-commercial terms and the NVIDIA Open Model License (Cosmos derivative). Do not use commercially.

Each runs/<phase>/<timestamp>/epochN/ is a ComfyUI-format LoRA — drop into ComfyUI/models/loras/.