開源模型與生成影像
Anima 3.8B v1.1 Integrates a Timestep-Aware Connector into the DiT, Allowing Dual Text Encoders to Be Offloaded Before Sampling
The community image model Anima 3.8B v1.1 uses Semantic Connector v2 to fuse Qwen3.5 4B semantics at every denoising step and packages the connector with a 52-block DiT in a single checkpoint. The new release simplifies deployment in ComfyUI and Forge Neo, but lacks standardized quality benchmarks, while full prompt encoding still incurs significant transient VRAM usage.
Anima 3.8B v1.1 updated its weights and workflows on September 1, replacing the previously separate progressive-cross adapter with adjustable strength with the timestep-aware Semantic Connector v2. This roughly 608M-parameter connector does not run only when generating conditioning; it remains inside the sampling model, extracting and injecting semantic features from Qwen3.5 4B at every denoising step according to the timestep. Its training strength is fixed at 1.0. As a result, developers no longer need to pair an adapter file or adjust a slider. The connector and the DiT—expanded to 52 blocks and approximately 3.8B parameters—have been merged into a single safetensors checkpoint.
Text conditioning still uses two paths: the native Qwen3 0.6B preserves Anima’s existing style alignment, while Qwen3.5 4B adds support for natural language, spatial relationships, and multi-character binding. Both text encoders are loaded only when the prompt changes and can be offloaded after conditioning is generated; they also do not need to run again when cached prompts are reused. This reduces VRAM usage during steady-state sampling, but the initial encoding or a prompt change still requires accommodating approximately 4.8GB of additional Qwen3.5 weights. The actual peak also depends on resolution, precision, the attention backend, and the frontend’s offloading strategy.
The accompanying ComfyUI nodes validate the native adapter hash recorded by the connector to prevent incompatible components from being packaged together, and safely reuse conditioning through a 64-entry registry. Forge Neo can also automatically identify v2 from safetensors metadata. The model’s author says the new version was trained on four A40 GPUs for approximately 140 hours and recommends starting at around one megapixel with 28 to 50 steps. However, the current evidence consists mainly of images showcased by the author, with no published FID, GenEval, or human blind-test results. Counting, hands, legible text, and crowded compositions remain listed as known weaknesses. The license also continues to follow the upstream Anima terms, so permitted model uses cannot be determined solely from the MIT license covering the accompanying code.