A fine-tuned model specializing in flat color, no lineart, deformed character art.
Note: This is my first fine-tuned model. While I'm pleased with this results, the quality may vary compared to established style LoRAs. I appreciate your patience and feedback as I continue to refine this approach. Your input helps me improve future versions!
This model excels at creating characters with a distinctive light deformation style - more expressive than realistic, yet more refined than super-deformed (SD/chibi). Perfect for characters ranging from 3.5 to 7 heads tall, capturing that sweet spot between youthful charm and mature appeal.
Key characteristics:
Clean flat colors (no lineart required)
Soft, natural deformation - ideal for stylized character design
Works beautifully with other character LoRAs to apply different character traits while keeping this style intact
Stable on deformed proportions that danbooru tags struggle to capture
Recommended usage: Combine with character LoRAs for maximum flexibility. The model's style strength is intentionally calibrated to enhance, not overwhelm, other adapters.
Training Details
For transparency and to help with troubleshooting, here's the training command I used with sd-scripts:
accelerate launch --num_cpu_threads_per_process 1 anima_train.py --pretrained_model_name_or_path=/path/to/animaOfficial_preview3Base.safetensors \
--qwen3=/path/to/qwen_3_06b_base.safetensors \
--vae=/path/to/qwen_image_vae.safetensors \
--dataset_config=/path/to/dataset_config_anima.toml \
--output_dir=/path/to/outputu \
--output_name=fwaunstp_ckpt_anima_preview3_v1 \
--save_model_as=safetensors \
--learning_rate=1e-6 \
--optimizer_type=AdamW8bit \
--lr_scheduler=constant \
--timestep_sampling=sigmoid \
--discrete_flow_shift=1.0 \
--max_train_epochs=8 \
--save_every_n_epochs=1 \
--mixed_precision=bf16 \
--gradient_checkpointing \
--cache_latents \
--vae_chunk_size=64 \
--vae_disable_cache \
--save_state \
--split_attn \
--attn_mode=torchThis model was trained using f32 precision (not BF16, which some other Anima Preview 3 models use). If you notice any issues or compatibility problems, please let me know—I'd appreciate feedback on the training parameters or the checkpoint format.
Description
FAQ
Comments (9)
I have a question. It's 7 GB... isn't it usually 4 GB?
I don't know why.
All I can say is that this is the result of training with sd-scripts.
As you mentioned, the other ANIMA models do seem to be 4GB.
I’ll research whether I can reduce it to 4GB by the next version.
@fwaunstp733 I get it now. Your model uses F32 precision, which is why it's 7GB, while the others use BF16, so they're 4GB.
I hope you don't delete the fp32 version. I want to see how it is compared with fp16 version.
Beautiful checkpoint.
I wanted to try this model in InvokeAI because the sample images are very cute, but I got an error when loading it.
Environment:
InvokeAI v6.13.0.rc2
Error:
RuntimeError: Checkpoint contains 3 unexpected keys. This may indicate a corrupted or incompatible checkpoint.
First 5 unexpected keys:
['pos_embedder.dim_spatial_range', 'pos_embedder.dim_temporal_range', 'pos_embedder.seq']
I checked this with Claude and ChatGPT, and the likely explanation is that this checkpoint contains extra positional embedding-related keys that InvokeAI’s Anima loader does not expect.
In particular, pos_embedder.dim_temporal_range looks like a temporal/3D positional embedding buffer from the Cosmos/Anima implementation. It may not mean the model itself is corrupted, but it looks like some extra training or implementation-specific buffers may have been saved into the checkpoint.
Other Anima Preview 3 based checkpoints work correctly in the same InvokeAI environment, so this issue may be specific to this checkpoint’s saved format.
Could you provide an InvokeAI-compatible or pruned version if possible?
Thank you for making such a lovely model.
Thank you so much for your interest in the model and for the detailed bug report! I really appreciate you taking the time to investigate and provide such thorough information.
A few clarifications on my end:
1. I've only tested this model in ComfyUI, so I don't have firsthand experience with InvokeAI's checkpoint loading process. I'm not entirely familiar with how the Anima loader expects checkpoint data to be structured.
2. I trained this model using sd-scripts, which is a standard training tool, but I'm not deeply knowledgeable about the underlying checkpoint mechanics or data structures involved. I'm honestly not sure why my checkpoint differs from other Anima Preview 3-based models in a way that causes this issue.
3. Another user did point out that my model uses f32 precision while other Anima models use BF16 (though I haven't independently verified this). This might be related to the checkpoint format differences, but I can't say for certain.
Unfortunately, I'm not currently familiar with how to convert or export the model in an InvokeAI-compatible format, so I can't provide a proper solution at this time.
I've added my training parameters to the model description for reference.
If you spot anything unusual in the training arguments or have suggestions for parameters that might improve InvokeAI compatibility, I'd be grateful to hear them.
Thank you again for the kind words about the sample images—I'm really glad they caught your eye!
This isn't really a discussion, but I learned from this model that you can summon Dragon Maids.
But Eldritch doesn't show up 😭