Disya/Anima-T5-Free-Base

Anima T5-Free Base

Anima T5-Free Base is a derivative of circlestone-labs/Anima that replaces Anima's original text-conditioning adapter with a T5-free conditioning frontend.

The Anima diffusion backbone is preserved. The original T5-dependent conditioning path and original llm_adapter are not used by this model.

This release is intended as a compatibility/parity base, not as a claim of improved prompt understanding over upstream Anima.

Base model: CircleStone Labs / Anima
ComfyUI node: ComfyUI-Anima-T5-Free
Support: Boosty

Architecture

The conditioning path is:

prompt
  ↓
Qwen3.5-2B-Base
  ↓
hidden states from layers 16 and 22
  ↓
T5-free row planner
  ├─ learned lexicon
  └─ lightweight OOV segmenter
  ↓
linear C0 projection
  ↓
residual C1 refiner
  ↓
512 × 1024 conditioning carrier
  ↓
Anima DiT

The model does not require:

T5 model weights
T5 tokenizer
Anima's original llm_adapter

The Anima DiT retains its existing 512 × 1024 conditioning interface. The carrier itself is generated entirely by the new Qwen-based frontend and does not contain T5 activations or require T5 at runtime.

Checkpoint layout

This repository follows a ComfyUI-style model layout:

diffusion_models/
    anima-t5free-fused-v0.1.safetensors

text_encoders/
    qwen_35_2b_base.safetensors

vae/
    qwen_image_vae.safetensors

The T5-free conditioner is already fused into the custom Anima diffusion checkpoint. Do not load the original Anima llm_adapter in addition to it.

ComfyUI support

This model requires a custom ComfyUI loader/node that understands the T5-free Anima checkpoint layout.

Stock Anima loaders expect the original conditioning architecture and are not compatible with this checkpoint.

Install the custom node, then place the files from this repository in the corresponding ComfyUI model directories.

Loading the fused diffusion checkpoint with an unmodified upstream Anima loader is unsupported.

What was changed

The upstream Anima diffusion model is retained, while its text-conditioning frontend is replaced.

The replacement frontend uses Qwen3.5-2B-Base representations and converts them into the conditioning representation expected by Anima's existing cross-attention layers.

The purpose of this model is to remove the runtime dependency on T5 while preserving the useful visual behavior of the original Anima model as closely as practical.

Validation

The release candidate was tested independently from the research tree.

The following checks passed:

• the release loader and the validated research implementation produced a bit-identical final latent for the same prompt, seed, sampler state and diffusion model;
• T5 model loading was blocked during the test and was never invoked;
• T5 tokenizer loading was blocked during the test and was never invoked;
• the original Anima llm_adapter was absent from the runtime path;
• normal prompts, a long tag-style prompt, a non-ASCII prompt, and a known difficult compositional prompt all completed full image generation without runtime failure.

The bit-identical validation above is between the validated T5-free research implementation and this packaged release implementation. It does not mean that this model is bit-identical to upstream Anima.

Intended use

This checkpoint is intended for experimentation and image generation with the Anima architecture without requiring the original T5-based conditioning path.

It can be used as a T5-free Anima base in workflows using the provided compatible loader.

Important limitations and risks

This is not a semantic upgrade

This release should not be interpreted as having better instruction following than upstream Anima.

It was designed primarily to replace the text-conditioning frontend while retaining useful behavior of the existing frozen Anima receiver.

Failures inherited from upstream Anima or from the compatibility mapping may remain.

Prompt adherence is imperfect

The model may fail on precise semantic requirements, including:

• exact object counts;
• negation such as no glasses, without ..., etc.;
• left/right and other spatial relations;
• ownership and attribute binding between multiple characters;
• complex multi-object composition;
• long chains of simultaneous requirements.

A visually plausible image does not imply that every part of the prompt was followed correctly.

Non-ASCII and multilingual prompts

The T5-free row planner was primarily validated on the text distribution used during development.

ASCII out-of-vocabulary words can use the learned internal segmenter.

Non-ASCII out-of-vocabulary text follows a more limited fallback path and has received substantially less validation.

Successful generation from a Russian or other non-English prompt should not be interpreted as proof of strong multilingual instruction following.

Fixed carrier size

The Anima receiver still consumes a fixed 512 × 1024 conditioning carrier.

Very long or unusually tokenized prompts can therefore hit the finite conditioning budget or behave differently from shorter prompts.

Exact compatibility with upstream Anima is not guaranteed

The diffusion backbone is based on Anima, but the text-conditioning frontend is different.

For the same prompt and seed, this model is not expected to reproduce the exact image produced by the original Anima text-conditioning stack.

Compatibility/parity here refers to preserving useful model behavior, not bitwise equivalence with upstream Anima.

Text encoder substitutions are unsupported

The conditioner was calibrated for its specified Qwen3.5-2B-Base text encoder representation.

Using a different Qwen model, an Instruct variant, a different hidden-state layout, aggressive text-encoder quantization, or another model revision may change the hidden-state distribution and degrade generation.

Such substitutions have not been validated.

Numerical/runtime variation

Different PyTorch, CUDA, attention-kernel, quantization, precision or ComfyUI environments can produce numerical differences.

The release was validated with the tested runtime configuration; exact reproducibility across arbitrary environments is not guaranteed.

LoRA and extension compatibility

Compatibility with third-party Anima LoRAs, ControlNets, training scripts, model merges and other extensions is not guaranteed.

Modifications targeting only unchanged Anima DiT layers may be structurally compatible, but this has not been universally tested.

Anything that expects or modifies the original Anima text adapter, text-conditioning keys, or original text-encoder path is incompatible unless explicitly adapted for this architecture.

Training compatibility

Training software written specifically for the upstream Anima state-dict layout may not recognize this checkpoint.

The conditioning architecture and checkpoint keys differ from upstream Anima, so custom loading/training support may be required.

OOV segmentation is approximate

The T5-free frontend reconstructs the carrier structure through a learned lexicon and segmentation mechanism rather than the original T5 tokenizer.

The reconstructed segmentation is not guaranteed to match the legacy conditioning structure for every possible prompt.

Rare words, unusual punctuation, mixed scripts, Unicode-heavy prompts and previously unseen strings may therefore behave differently.

Image quality and generation failures

Like other diffusion models, this checkpoint can produce:

• malformed anatomy;
• incorrect hands or faces;
• duplicate or missing objects;
• incorrect text;
• inconsistent perspective;
• implausible object relationships;
• unexpected characters or attributes;
• visual artifacts.

The model can also ignore part of a prompt while still producing an aesthetically coherent image.

Bias and dataset limitations

The behavior of this model inherits biases and limitations from the upstream Anima model, its training data, the Qwen text encoder, and the data used to construct the replacement conditioning frontend.

Outputs may reproduce stereotypes, dataset imbalances, common visual tropes, artist/style biases, or other undesirable correlations.

Safety

This model does not provide a guarantee that generated content is safe, accurate, appropriate, non-infringing, or suitable for a particular use.

Users are responsible for reviewing generated outputs and complying with applicable laws, platform rules and licenses.

Reproducibility notes

For meaningful comparisons, keep the following fixed:

model revision
text encoder revision
VAE
sampler
number of steps
CFG
seed / initial noise
precision
ComfyUI/custom-node version

Changing any of these may materially affect the result.

Support

If you find this project useful and want to support the work:

Support me on Boosty

This is optional support only. There are no exclusive model files, early-access checkpoints, or gated content attached to the subscription.

Attribution

This model is derived from:

CircleStone Labs — Anima

https://huggingface.co/circlestone-labs/Anima

Anima itself is derived from / related to the architecture lineage shown in the Hugging Face model tree.

This repository contains a modified text-conditioning architecture and does not claim authorship of the original Anima diffusion model.

License

The CircleStone model components are licensed under the CircleStone Non-Commercial License.

See LICENSE.md for the license text included with this repository.

Use of this derivative model remains subject to the applicable upstream license terms.

The license and attribution notices of third-party components must also be respected.

This model card is descriptive and is not legal advice.