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/.
