Anima-Telescopa

Telescopa is a full matrix LoKr (basically poor man finetuning) of Anima Base v1.0 by CircleStone Labs, trained using the LyCoRIS full matrix LoKr method (paper).

This was my first attempt at fine-tuning an image model on a dataset above 10k samples. Alongside general practice, the run was also used to:

• Test whether a carefully hand-curated aesthetic dataset could improve Anima's background generation quality

• Compare full matrix LoKr training against a true full finetune on the same settings

Compared to other Anima fine-tunes such as Tdrussel's Aes B, Motimalu's KirazuriV4, duongve's AnimaYume, and the Silvermoon mixes,ect,.... Telescopa is perhaps a lil worse on fine detail, stability, and knowledge/style retention — but it shows slightly stronger background composition and detail than the base model and most contemporaries. For a first larger-scale training run, I'm happy with the result and want to release it anyway as a open source showcase and as a lesson/milestone for myself.

This model is produced independently, as a hobbyist project, with no external funding. The model page contain 3 model variant, the ideal epoch 10 LoKR, a earlier and more unstable epoch 3 that also doesnt have as much style bias or potential knowledge forgetting, and a failed epoch 10 full finetune.

For in-depth details of training and tooling, see:

• Anima-Telescopa Training Diary

Training Details Summary

Trainer: diffusion-pipe-bluvoll-fork

Training device: x2 NVIDIA RTX™ A4000 16GB VRAM provided by Astromahdi gpu-garden

Total training time: ~3 days or ~68 hours (more accurately about 2 days at best, I screwed up and had multiple failed runs/restarts)

Total samples seen(unbatched steps): ~100,0000 samples

Training resolutions:

• 1024^2

• 1280^2

Training run ( initial 3 epoch run into a further 7 epoch 2nd run)

See: TelescopaLOKR training config

• Samples seen(unbatched steps): ~30,000 + ~70,000 samples

• Learning Rate: 3e-6

• Learning Rate Scheduler: Constant with Warmups

• LLM Adaptor Learning Rate: Disabled

• Precision: Full BF16

• Optimizer: AdamW8bit with Kahan Summation, utilizing Kahan summation to prevent precision roundoff errors during pure BF16 weight

• Weight Decay: 0.01

• Timestep Sampling Strategy: Logit-Normal with Shift 4 and Sigmoid Scale=1.3

Additional Features

• Tag Dropout: 10%

• Caption Dropout: 5%

• Mixed caption with ratio of 25/25/25/25

• Tag Shuffle

• Enable Full Matrix LoKR with: dim/rank = 16 alpha = 16 dtype = 'bfloat16' dropout = 0.0 factor = 2 (~800mb file size)

Dataset

Collected through various online sources, mainly danbooru and Akanyan personal collection of anime screencap, scattered throughout his account.

I fully open source the dataset as a gesture of goodwill towards general open source, which I've depended on for all my times of using AI/machine learning. And also because open sourced datasets are rare and doing this gain me more credibility lol: RicemanT/Anime-Background-Finetuning-V1.1 · Datasets at Hugging Face

Installing and running

Workflow:

Reference the anima base instructions. The model is natively supported in ComfyUI. The above image contains a workflow; you can open it in ComfyUI or drag-and-drop to get the workflow.

Note: This is just my personal simple workflow, you can realistically use whatever you want as long as the setup work well.

The model files go in their respective folders inside your model directory:

• The failed epoch 10 finetune goes in ComfyUI/models/diffusion_models

• The actual successful LOKR models (which is a LoRA variant at its core) goes in ComfyUI/models/loras

• qwen_3_06b_base.safetensors goes in ComfyUI/models/text_encoders

• qwen_image_vae.safetensors goes in ComfyUI/models/vae

Generation Settings

These are the settings used for the sample images above (ComfyUI, ModelSamplingAuraFlow node):

• Shift: 5.0

• Steps: 28

• CFG: 4.0

• Sampler: er_sde

• Scheduler: simple

These are just my usual settings — feel free to experiment and really go buck wild with it, euler a at same steps and sa_solver_pece at a lower 15–20 steps also recommended.

Prompting

Like the base model, this model is trained on booru-style tags and natural language captions.

Quality and Aesthetic tags + Negatives (same as base model)

masterpiece, best quality, score_7, safe

worst quality, low quality, score_1, score_2, score_3, artist name, blurry, jpeg artifacts

or anything you prefer.

Known Limitations & Issues:

• Some instability is noticeable when using short prompts, this behavior is similar to the base model and several other finetunes.

• Slightly degraded text performance from some of my initial testing and personal experience.

• Mentioned above, artist tags are not as faithful as I hope, the model seems to diverge a bit and have a orange bias.

There are probably more but I haven't got enough uses on this model myself to completely explore. If you guys encounter major quirks or problems, do tell in the comments, I do actively read them and knowing there are problems means I can learn from it and solve it in future projects.

Recognitions

Beeg thanks to:

• CircleStone Labs — for Anima and the original diffusion-pipe trainer

• Motimalu — Inspiration behind finetuning practices and configs, his diary was such a huge help for a completely beginner like me to get an idea behind how finetuning a model looks like

• Bluvoll — diffusion-pipe fork used for training, and general training advice

• Anzhc — general training advice

• Akanyan — for the goated collection of anime screencap from 9 years ago

• Astromahdi — compute via gpu.garden

• KohakuBlueleaf — LyCoRIS / LoKr library

• animetimm/DeepGHS — tagging model

• Model description format inspired by Motimalu's

License

This model is released under the same license as the base model.

See the base model for details of the CircleStone Labs Non-Commercial License.

Built on NVIDIA Cosmos